Amplifying Human Capabilities in Prostate Cancer Diagnosis: An Empirical Study of Current Practices and AI Potentials in Radiology

This paper examines the potential of Human-Centered AI (HCAI) solutions to support radiologists in diagnosing prostate cancer. Prostate cancer is one of the most prevalent and increasing cancers among men. The scarcity of radiologists raises concerns about their ability to address the growing demand for prostate cancer diagnosis, leading to a significant surge in the workload of radiologists. Drawing on an HCAI approach, we sought to understand the current practices concerning radiologists’ work on detecting and diagnosing prostate cancer, as well as the challenges they face. The findings from our empirical studies point toward the potential that AI has to expedite informed decision-making and enhance accuracy, efficiency, and consistency. This is particularly beneficial for collaborative prostate cancer diagnosis processes. We discuss these results and introduce design recommendations and HCAI concepts for the domain of prostate cancer diagnosis, with the aim of amplifying the professional capabilities of radiologists.


INTRODUCTION
Every year, thousands of individuals are diagnosed with cancer.On a global scale, there were 1,276,106 recorded instances of prostate cancer in 2018, making it the second most common cancer in men worldwide [22].In Germany, where this study was conducted, prostate cancer stands as the most prevalent cancer among men, with approximately 68,579 cases diagnosed in 2019, predicted to reach 70,100 new cases in 2022 [38].Given demographic trends, a projected 23% increase in new cancer cases is anticipated in Germany from 2015 to 2030 [38].
This dynamic contradicts the limitation of available radiologists, approximately 9,535 radiologists, constituting merely 2.3% of the total count of medical doctors [23], who are critical resources in the diagnosis of prostate cancer.This scarcity raises concerns about the adequacy of specialized medical professionals to address the growing healthcare demands [60].These statistics underscore the gravity of prostate cancer, the shortage of radiologists, and their overall workload.Hence, it is logical to explore potential solutions that can support these experts in their roles, promoting enhanced productivity and efciency.
Incorporating AI into radiology could potentially assist physicians by facilitating easy and rapid informed decision-making, thereby aiding in overcoming bottlenecks.AI systems possess the potential to boost diagnostics and treatment planning through the utilization of extensive and varied research and personal healthcare data [6,70].AI's ability to swiftly process substantial volumes of data surpasses human capabilities [91], ofering error reduction [96], ofering medical professionals a preliminary diagnosis, enhancing the potential for refning fnal assessments [8].Consequently, AI could act as an assistant, fostering collaborative diagnosis through shared control.This concept embodies a nascent interaction framework where both humans and AI collaboratively oversee system management [31].Radiology, with its highly data-intensive nature of imaging procedures, is particularly suited for AI applications.Radiologists, as key stakeholders in imaging modalities, can harness AI's potential for improved accuracy, efciency, and consistency [54,104,106].Inevitably, AI holds signifcant promise in enhancing prostate cancer diagnosis for radiologists.
Despite substantial advancements in AI within healthcare, only a few AI systems have been efectively integrated into medical practice [89].This can be attributed to the oversight of involving humans in the development process [2].Understanding people's practices and how technological innovations can be integrated into those practices or how they can foster the establishment of new (and hopefully enhanced) ones is key [113].Therefore, to shift AI from being merely technological to becoming socially integrated, we contend that adopting an approach combining Human-Centered AI (HCAI) and practice-centered design is imperative [102,113].
Hence, in this paper, we will argue that the potential of AI, facilitated by HCAI, lies in the ability to generate healthcare outcomes that are not only more accurate but also more efcient compared to conventional physician-centered methods.We adopt a practicecentered design process with a strong focus on involving practitioners in the design process to fully understand current practices and seamlessly integrate the new design into it [112], as is common in HCAI initiatives.
Our primary contributions aim to address the following Research Questions (RQ): • RQ1: What are the unresolved challenges faced by radiologists in current practices concerning prostate cancer diagnosis?• RQ2: What potential does AI hold in enhancing radiologists' current practices for prostate cancer diagnosis?
To answer our RQs, we have engaged in extensive qualitative feld research carried out in four Radiology Center (RC) in Germany.Data collection was mainly driven by contextual inquiry [11] and indepth semi-structured interviews.Through contextual inquiry, we were able to observe and interact with radiologists, who work with prostate cancer patients, as they analyze and diagnose potential cases.The collected data was analyzed thematically using Braun and Clarke's approach [21].This helped us outline the design space for technologies in prostate cancer diagnosis and explore AI's potential to address identifed design opportunities.
We present a three-fold contribution: an empirical study on radiologist practices in diagnosing prostate cancer; an informed discussion on how AI can enhance decision-making; and design recommendations and HCAI concepts for prostate cancer diagnosis.Our suggestions span attitudes toward the use of AI in radiology, AIdriven visualizations, automation for diferent tasks, amplifcation of human capabilities through hybrid intelligence, and AI-based decision support tools, aiming to support and enhance radiologists' performance in prostate cancer diagnosis.
The remainder of this paper is organized as follows: we provide an overview of related research on AI in radiology and prostate cancer diagnosis in section 2. We move on to describe our methodological approach in section 3, and then go on to explore current practices and challenges in prostate cancer diagnosis in section 4, thus addressing RQ1.Subsequently, in section 5, we address RQ2 by proposing a conceptual solution to support the challenges through AI, informed by our empirical results and literature insights.In section 6, we discuss how our ideas bring the state of the art forward to support radiologists in identifying and diagnosing prostate cancer.Finally, we conclude by refecting on our empirical study referring to our RQs.

RELATED WORK
In this section, we explore existing literature that underpins our study, immersing ourselves in the realm of AI within the context of radiology.Here, we discover how technological innovations have enhanced diagnosis accuracy and efciency, refecting AI's importance as a key work infrastructure across domains [29].Next, our focus turns to prostate cancer diagnosis, where we examine how AI is currently being used.We also discuss why these advances do not always translate seamlessly into real-world medical practice, emphasizing the importance of keeping humans at the heart of AI development lifespans [49,61], including developers and researchers in charge of building AI solutions [101], and end-users who will actually use these solutions [89].By navigating through these intertwined thematic trajectories, we outline the intersections among AI, radiology, prostate cancer diagnosis, and human-centered design.

AI in Radiology
Healthcare, especially radiology and cancer diagnosis, can greatly beneft from the use of AI by enhancing decision-making through prognostic information and potentially accelerating cancer imaging and screening.AI systems can provide greater diagnostic performance over humans in less time but without exhaustion, as well as mitigate human faws and limits, reducing preventable errors [53,76,103,107,121].There are several studies in similar healthcare settings about attitudes and potentials toward AI in pathology to increase efciency and accuracy [69,84].Also in radiology, few studies explored the potential of AI as a support tool.For instance, Griesshaber et al., [47] examined the integration of AI-based systems in radiology from the radiologists' perspective.One type of clinical decision support system used in radiology is the Computer-Aided Detection (CAD) system which "supports the detection of lesions, classifcation of the disease, diagnostic decision making, and predictions regarding treatment, but the fnal decision is still made by radiologists" [47].Interestingly, Oktay et al.'s research on radiotherapy and cancer care aligns with our focus on workfow optimization [88].
Several studies can be found for breast cancer aiming to integrate and evaluate support tools following current practices [50,51].Bi et al., [12] conducted a study in that context addressing the use of CAD systems as decision-making tools to detect missed tumors or microcalcifcations in mammography, and detect early signs of breast cancer.AI models are helpful for such CAD systems because they can analyze the Magnetic Resonance Imaging (MRI) scans following the nuances of existing human clinical practices [26].Decision-making processes with the use of AI in radiology were also studied by Cabitza et al., [24], which reveals that suitable interaction protocols allow for higher precision in decision-making that surpasses the performance of individual diagnosticians, whether radiologists or AI.

AI in Prostate Cancer Diagnosis
There exist a variety of studies on the use of AI for prostate cancer diagnosis.Many examples include Convolutional Neural Network (CNN) applications in medical imaging for the detection and segmentation of prostate cancer [30,43,56,71,82,120,122].The focus of these studies is on the use of neural networks for image classifcation, analysis, and segmentation, using MRI images, and improving the CNN and training approaches of these algorithms.For instance, in 2019, Huang et al., proposed a CNN architecture for efciently segmenting prostate cancer tissue in Multiparametric MRI (mpMRI) images.They rescaled and merged the images and used a CNN architecture for training, achieving high accuracy in the segmentation [56].Additionally, Ghavami et al., measured the accuracy of six distinct CNNs in segmenting the prostate in 3D-patient T2-Weighted (T2W) MRI scans.These diferences in CNN approaches should be considered when selecting between diferent network architectures, along with reporting the segmentation accuracy [43].
There are more studies in which radiomics features or Machine Learning (ML) or Deep Learning (DL) techniques have been tested for the detection or classifcation of prostate cancer [17,27,67,111].One example is a study by Wang et al., showing that "ML analysis of MR radiomics can help improve the performance of Prostate Imaging Reporting and Data System (PI-RADS) in clinically relevant Prostate cancer" [111].Ginsburg et al., demonstrated that radiomics features can be efective for prostate cancer detection, but the usefulness difers in diferent zones, suggesting that "decision support tools for evaluating prostate MRI exams should take into account diferences between Peripheral Zone (PZ) and Transitional Zone (TZ) tumors" [44].In 2018, Bonekamp et al., conducted a study comparing Radiomic ML (RML), mean Apparent Difusion Coefcient (ADC), and radiologist assessment to characterize prostate lesions found in routine clinical MRI interpretation [17].The study focused on lesion characterization rather than detection because lesions were manually identifed.
Though AI has reached the clinical implementation stage in healthcare, the practical efectiveness of AI algorithms for screening certain diseases still remains unsatisfactory [75].Several pieces of research have also revealed how medical practitioners overlooked suggestions identifed by AI [35,86] and, also missed anomalies that AI failed to identify [63] because of distrust or overtrust of AI, respectively.On the one hand, AI helped physicians identify breast cancer quicker and better [68], but on the other hand, certain AI systems fared badly and might unwittingly cause more damage than good if used to infuence treatment decisions [114].
Other studies look in particular at the needs of pathologists and how to increase diagnostic utility and user trust [25].Verma et al., suggest that "to develop physicians' trust in AI and its wider acceptability in clinical oncology, designers would have to address the ethical and liability concerns about the use of AI systems" [110].Due to the subjectivity inherent in cancer grading, it becomes important for practitioners to uncover the reasons behind AI decisions to trust the system's result as the clinical background of an AI can be opaque in contrast to practitioners being aware of their colleagues' medical knowledge and experience, involved in the diagnosis [26].Hence, physicians are required to know the inputs and context accessible to the algorithm, and the basics of how it analyzes those inputs to make a decision.Their tendency to relate AI behavior to human schemas suggests an AI primer could be useful by comparing AI processes to human clinical processes [26].For this purpose, it becomes imperative to focus on three aspects of decision-making processes: frst, the design of AI decision support systems; second, the human-machine interaction they entail; and third, the outcome of the decision-making process for clinical treatment [14].

Human-Centered AI (HCAI)
Along with recent research, initiatives like the European Humane AI Project [80] and the Stanford Institute for HCAI [40] focus on the human aspects of AI working toward HCAI, as well as making AI interactive and explainable.These initiatives investigate human factors such as trust, fairness, and transparency which are signifcant for the acceptance of AI applications [99].
Other studies relevant to our research [42,105,119] consider HCAI design, ideation, early prototyping, and problem formulation.Thieme et al., suggest that addressing the challenge and ambition of AI in clinical settings requires specifc attention to design issues, particularly those concerning time constraints in treatment [105].Yildirim et al., consider the challenges of AI for user experience design and suggest an improved ideation approach for overcoming the tension between user-centric and technology-centric design approaches [119].
When it comes to HCAI, performance and explainability are often seen as trade-ofs.For instance, Kelly et al., [64] present results suggesting that the more explainable an AI model, the lower its performance.Nevertheless, there has been a growing movement toward the argument that enhanced explainability can potentially support users to make sense of results and assist them with decisionmaking [48].Researchers have, therefore, argued for the strong involvement of end-users in the design of eXplainable AI (XAI) interfaces [9], elaborating implications for design, such as the design of interfaces to provide descriptions of algorithmic decisions; work on various layers of rationalization; and make access to complete data origin and related boundaries viable [101,117].
Furthermore, researchers from the community have argued that engaging users in the design process and especially in design decisions -according to the participatory design premises [13,20] -can contribute toward more ethical, adaptable, and useful AI systems, prepared to face adverse unexpected efects [16].XAI techniques can arguably contribute toward accelerating the adoption of AI solutions in medicine, by supporting increased transparency in the generation of results which may have a positive impact on the trust relationship between users and the AI system in question.This transparency was found to be sometimes decisive in the adoption of such systems [4].
Recent research highlights the signifcance of visualization in achieving transparency.Lee [72], for instance, discusses how the decisions of AI systems can be made more understandable and transparent through visualizations that lead to a fair, responsible perception.Visualization, hence, can be considered the frst step of explainability.At the same time, explainability may mitigate some distrust or overtrust issues which can make the use of such systems in healthcare problematic.
This article seeks to address the existing gap by conducting empirical research to identify the real challenges in current radiological practices for prostate cancer diagnosis, which can be alleviated through AI integration.Our focus is on introducing HCAI solutions with a qualitative approach, to provide better improvements [39] in terms of enhanced accuracy, efciency, and consistency, particularly through the use of visualizations for radiologists.

METHODOLOGY
We structured our research process following the premises of HCAI and practice-centered design, which essentially means keeping people and their practices at the heart of AI solutions.Through this human-centered approach [37], we try to understand and prioritize end-user needs and pain points.We collaborated with diverse stakeholders, including HCI researchers, AI developers, medical experts such as radiologists, and Medical Technical Radiological Assistant (MTRA)s to address our RQs.Our main focus was on radiologists, recruited primarily in cooperation with the German Radiological Society 1 .
To explore the present context of use within the real user environment, we conducted qualitative feld research, employing a combination of contextual inquiries and in-depth semi-structured interviews.We opted for contextual inquiry as our main method to gain a comprehensive understanding of the context and directly observe users in their natural work settings [59].To capture crucial insights beyond interviews, we immersed ourselves in the radiologists' work environment, acknowledging that established routines and implicit practices may not surface during interviews.These inquiries not only let us observe their actions but also led to insightful discussions, ensuring a thorough exploration of their current practices .Additionally, we opted to use in-depth semi-structured interviews to delve deeply into the radiologists' insights regarding the diagnosis process.The semi-structured interview is acknowledged as a potent tool in medical research for understanding individuals' beliefs, experiences, and perspectives [36].
Our study design underwent ethical review by our institution's IRB.Participants received study information, privacy, and data security details, and signed a consent form before participating.
To gather empirical data, we conducted contextual inquiries across three distinct RC located throughout Germany.These inquiries involved a blend of observation, conversation, and interactive exploration, each session spanning three to fve hours in each RC.Initially, our feldwork commenced in RC01 (visited by one researcher) and extended over four months, engaging with radiologists P01 and P02, along with MTRAs.Subsequently, we extended our empirical data collection to two additional RCs (both visited by two researchers) over two months, collaborating with P03 in RC02, and P04 in RC03, as well as the MTRAs from both centers.Our investigation encompassed a total of eight contextual inquiry sessions across RC01, RC02, and RC03.This groundwork was further enriched through four semi-structured interviews with radiologists 1 Deutsche Röntgengesellschaft: www.drg.deP01, P02, P03, and P04.Furthermore, an interview with radiologist P05 was included in our study, focusing on his experiences in RC04, which we regrettably couldn't visit in person.The interviews, ranging from 40 to 80 minutes in duration, were conducted online via Zoom2 .For the data collection, both audio recordings and written notes were employed for documentation purposes.To safeguard participants' identities and confdential information, all the transcriptions and translations were pseudonymized throughout the research process.
For a comprehensive analysis of our collected data, we employed a thematic analysis approach [21].This phase involved the collaborative eforts of our interdisciplinary project team [19], comprising three researchers and two to three developers.Immediately following the data collection phase, we embarked as a team on a meticulous process of data familiarization.This phase encompassed revisiting and replaying signifcant excerpts from the audio recordings of both our observations and interviews.Additionally, we reviewed our handwritten notes taken during the sessions.This thorough examination allowed us to identify prevalent patterns, challenges, and emerging user needs within the context of prostate cancer diagnosis.
The data were systematically coded in a shared document among the team members, with each code capturing a specifc aspect or insight gleaned from the empirical study.To decode the data, two researchers used Miro board 3 , a collaborative online whiteboard tool, to create an afnity diagram, visually organizing and connecting related information.This diagramming process served as a vital bridge between raw data and actionable design concepts, by "keeping design teams grounded in data as they design" [83].We discussed new insights through regular meetings again with the whole team between researchers and developers.The identifed themes and their implications form the basis for the discussions presented in the subsequent sections 4 and 5, providing a comprehensive exploration of our fndings and their relevance to the feld of prostate cancer diagnosis.The research process model of our empirical study is depicted in Figure 1.

CURRENT PRACTICES AND CHALLENGES IN PROSTATE CANCER DIAGNOSIS
In this section, we address RQ1, "What are the unresolved challenges faced by radiologists in current practices concerning prostate cancer diagnosis?".The analysis of data reveals a multifaceted understanding of radiologists' current work practices.This includes a comprehensive overview of the diagnosis procedure, diverse approaches employed, the use of diferent artifacts and systems, and both internal and external communications.We identifed numerous challenges faced by radiologists in their radiology work during the prostate cancer diagnosis process.

Current Diagnosis Procedure
Our main focus was on the responsibilities of radiologists in diagnosing prostate cancer.In this context, radiologists aim to establish diagnoses and enhance the identifcation of clinically signifcant prostate cancers before they progress to a point requiring treatment.In the initial phase, radiologists engage with patients before the examination to elucidate the procedure and gather relevant medical history.This includes details such as Prostate Specifc Antigen (PSA) values, pre-diagnosis information (e.g., biopsy results), Gleason score, medication history, and factors like the presence of a pacemaker or sports activities (e.g., cycling, which can infuence PSA values).This patient interaction is crucial to collect pertinent meta-information as an initial assessment.P04 emphasized the necessity of at least one PSA value, preferably multiple measurements.He enumerates several factors that could lead to an increased PSA level: "One is a tumor, another could be cycling, and yet another could be prostate infammation." Signifcantly, PSA values play a crucial role in determining the necessity of an MRI examination.Urologists are well aware of this protocol, often referring patients to radiologists when sustained elevated PSA values are observed.This holds particular signifcance for patients covered by public health insurance, as they bear the expenses themselves, and unnecessary costs can be averted.[73], an MRI of the prostate should be performed multiparametrically while following the protocols from the latest PI-RADS (v2.1 [108]).That mpMRI exam of the prostate should include the following sequences: T1-Weighted (T1W), T2W, Difusion-Weighted Imaging (DWI), and Dynamic Contrast-Enhanced (DCE) [108].Our observations reveal that all RC follow that guideline.
During the patient consultation, P04 highlights the importance of MRI, especially when there is an elevated PSA value.MRI allows for precise biopsy targeting, reducing the need for blind procedures and multiple punctures.P04 noted: "The MRI does something good [...].If we see a lesion, we can tell the colleagues [urologists] that they have to specifcally point there [specifc location(s)].And that's fne." Due to lingering mistrust in MRI accuracy [17,18], biopsies are still formally recommended; however, the goal is to minimize them due to their potential discomfort for patients [5].
After the initial meeting, patients are handed over to the MTRA, who is typically responsible for conducting MRI exams, reviewing the images, and making adjustments if necessary.Radiologists can access the images through the Picture Archiving and Communication System (PACS).

Image Interpretation. The primary role of a radiologist is to interpret imaging data generated by MRI scanners with precision, aiming to accurately detect and assess the presence of lesions.
Radiologists employ PI-RADS as a classifcation scheme to discern the presence of prostate cancer [73].This scheme evaluates the probability that a suspicious region indicates an acute malignancy.It assigns a categorization to abnormal prostate lesions on a 5-point scale, ranging from PI-RADS 1 (indicating minimal likelihood of being a signifcant clinical carcinoma) to PI-RADS 5 (signifying a high likelihood of being a major clinical carcinoma) [108].
From the image interpretation and certain related calculations, a PI-RADS score is derived.Section 4.2 elaborates on this process to enhance a deeper understanding, considering its signifcance.
4.1.4Reporting.We observed, that all radiologists used a dictation system (see section 4.4) to generate or modify their reports.Rather than having data transfer occur automatically, the radiologist needs to verbally dictate all measurements and calculations derived from MRI images during the diagnosis.P01 resorts to copy-pasting reports from other patients as templates.P02 uses text modules for the report, given the time-intensive nature of prostate examinations.Furthermore, P04 follows structured, standardized reporting, adhering to the template of the German Radiological Society by copying the standard report.
Besides the MRI images, radiologists also get patients' fles from the administrations containing, e.g., the referral, patient information, note sheets with handwritten notes (e.g., PSA value, previous examination results, etc.), which they use for the report.The report includes patient metadata, pre-diagnosis details (including the PSA values), and the fnal diagnosis comprising the PI-RADS score, calculation outcomes, and potential recommendations for subsequent actions.The fnal report also comprises a prostate sector map depicting the lesions (see section 4.2).
4.1.5Verification.The ffth phase is optional as communication with fellow radiologists is not a routine aspect of the daily radiological practice.However, radiologists engage in phone communication with other involved doctors whenever necessary to discuss special cases.Radiologists often request supplementary tests to validate their fndings, underscoring the signifcance of needing verifcation of their diagnoses.
Prostate cancer is prevalent among men [22], yet a signifcant proportion of cases are regarded as clinically insignifcant, characterized by very low metastatic potential.These are often cancers that men will die "with" rather than die "from" (P03).P04 emphasizes this point by saying that 30-40% of men get prostate cancer in their lifetime, but many of them don't even notice it.These tumors sometimes grow extremely slowly, so that people can live with them and do not actually die because of them.Hence, radiologists need to verify their prostate cancer diagnoses due to the tumors' slow growth and potential variability in signifcance.This careful assessment helps distinguish between malignant and benign cases and ensures appropriate medical decisions are made, avoiding unnecessary treatments and possible complications.

MRI Interpretation using PI-RADS Classifcation
T1W images are used primarily to determine the presence of hemorrhage and to identify the outline of the gland.T2W images derived from diferent orientations (axial, coronal, and sagittal) give an overview of the prostate zonal anatomy, such as PZ and TZ.About 70-75% of prostate cancers are found in the PZ and 20-30% in the TZ (P01, [108]).T2W is the dominant sequence for TZ while DWI is for PZ (P03).DWI is comprised of two sets of images for analysis, the high B-value (b-value) images, and an ADC map with diferent low b-values.Without exception, DWI fndings should be correlated with T1W, T2W, and DCE [108], allowing radiologists to assess specifc lesions by comparing the various images.DCE is the capture of fast T1W gradient echo images before, during, and after the use of a gadolinium-based contrast agent, which aids in the identifcation of small lesions.Not every lesion can be rightly characterizable in every picture.Hence, it is important to look at all these diferent pictures because they are generated from diferent perspectives to determine a PI-RADS score.During our observations, we noticed that radiologists were going back and forth, comparing the images to make sure that they didn't miss out on details, and interrogating the given images more closely to detect lesions.
Prostate MRI interpretation is complex, leading radiologists to rely on PI-RADS guidelines to score lesions for the likelihood of clinically signifcant prostate cancer [90,108].Radiologists frst assess the lesion location (PZ or TZ) and then the dominant sequence, applying PI-RADS criteria to score each location in the prostate based on dominant sequence parameters.
Furthermore, radiologists measure manually the gland by width, height, and length using the ruler tool within the diagnostic software by dragging the digital distance meter on the MRI to get the measurements from three perspectives (x-, y-, z-axis), and calculate prostate volume using the ellipsoid formula (W×H×L×0.52).We observed them recording the measurements on a sheet of paper.They then searched through the patient's documents to fnd the PSA value, which sometimes was noted on a separate sticky note by the MTRAs or the administration (RC01).Afterward, they calculated the PSA density by using the PSA value divided by the prostate volume.Besides P05, all radiologists primarily utilize the calculator app on their personal mobile phones for these calculations.
The results of prostate volume and PSA density are then recorded manually in the report, a process that can be error-prone.During our feldwork, we observed that P01 made a typing error in the calculator, resulting in an incorrect outcome.Due to his extensive experience, the doctor quickly identifed the error by cross-referencing the results with the MRI image.Recognizing the mistake, he recalculated to rectify the error by spending extra time.
According to PI-RADS, a standardized prostate sector map is used in which identifed lesions are marked.The exact localization of the lesions is essential [55,97] e.g., for a biopsy.We observed that radiologists use prostate sector maps to show the location of the lesions for their fnal diagnosis reports.Radiologists use pre-printed maps and draw on them by hand to highlight the locations for referring doctors which takes up extra time.P04 complained about having to manually draw lesions into the paper-based prostate sector map.
However, radiologists have shared their negative views about the PI-RADS scheme too.P03 mentioned that PI-RADS gives the impression that the diagnostic process is straightforward.He notes that the accuracy of prostate volume calculation can vary among radiologists due to the manual distance measurement.This variance can potentially lead to issues, as PI-RADS imposes stringent limits for critical values.P03 further emphasized that all these classifcation schemes like PI-RADS neglect that there is a human who assigns the score to MRI images.He referred to the problem by saying: "Even if all the criteria of PI-RADS classifcation are met, radiologists end up overestimating 80% of the cases and underestimating 5% of the cases, although it is rare for signifcant carcinomas to go unnoticed."

Diverse Approaches to Prostate Diagnosis
Variations in diagnosis approaches emerged as a notable aspect across diferent RCs, with distinctions arising between individual diagnosis and double diagnosis methods by radiologists.
At RC01, radiologists follow a double analysis policy for prostate exams, improving anomaly detection and diagnostic accuracy by having two radiologists assess each case.After P01 completed his analysis, it was subsequently verifed by P02, and with the fnal decision, the report was generated.In contrast, RC02 and RC03 adopt a single analysis approach, with P03 and P04 conducting individual diagnoses for each case alone.However, P04 mentioned that he occasionally engaged in discussions with peers via phone after sharing images to seek feedback on complex cases.Other radiologists held meetings with urologists to deliberate on critical cases (P03, P05).
In RC01, we observed some decision conficts between the two radiologists, which could lead to diferences for further steps.These exchanges hold signifcant value prompting them to invest more time in image analysis to reach a mutually agreed conclusion.However, we observed that the second radiologist often has limited time for image analysis due to time constraints.Consequently, the frst radiologist provides a summary to the second, potentially leading to a biased perspective as the second radiologist might not thoroughly examine the images.Moreover, if the second radiologist is unavailable, there is a delay in completing diagnoses requiring additional efort for each case (P01).
At RC03 it is typically and ofcially recorded that they perform a double diagnosis.However, in practice, the radiologist sends the MRI images to colleagues from another clinic only for more complex or critical cases and asks for a second opinion (P04).P04 further suggested that, as a general practice, approximately 10% of all fndings should be subject to blind re-evaluation by others, without any prior agreement."The fndings would then be so much better if they were looked at again in pairs" (P04).
During the interview with P05, we learned that their RC does not practice double diagnoses, despite having fve radiologists, who are experienced in prostate MRIs.In the past, P05 engaged in double diagnoses as a form of supervision, especially for new colleagues.However, as P05 explained, radiologists quickly develop a routine and become profcient in image analysis terminology.In cases of uncertainty or unclear fndings, colleagues collaborate, reviewing and discussing challenging situations before fnalizing the report.
Radiologists occasionally review their own diagnosis for a second time, ideally on the following day, to gain a fresh perspective and enhance the accuracy of their assessment.This approach mitigates the infuence of hasty judgments that might occur during the initial diagnosis, thus enhancing the credibility of their evaluation when obtaining a second opinion isn't feasible.In RC03, P04 employs a similar strategy by revisiting fndings the next day if a second opinion isn't available.He said, "If you have time, you can do it at your leisure.It's just better to look at it a second time, without having to worry about it.I just think that's important".In RC01 it was mentioned that one gets tired, especially in the evening from the many image reviews, so concentration is lost.Therefore, the radiologists check the images again the next day before sending out the fnal report (P01).

Current Artifacts and Systems
Throughout our observations, a recurring theme emerged: each RC employs distinct workfows, along with specialized hardware and software solutions tailored to streamline the diagnosis process.We highlight key artifacts and systems common across all centers, which collectively contribute to the seamless workfow of prostate cancer diagnosis.
All the RCs use a similar infrastructure for the overall diagnostic procedure.They have standard systems, such as the Hospital Information System (HIS), Radiology Information System (RIS), and PACS.HIS is mainly responsible for administrative tasks.RIS is used by radiologists to edit, report, and store radiology diagnoses.PACS is a medical imaging technology used for transferring, archiving, and accessing medical images from various modalities such as MRI by PACS Server, and displaying images by using a medical imaging software, Digital Imaging and Communications in Medicine (DI-COM) viewer through PACS Clients.DICOM is an open standard for storing and exchanging information in medical image data management 4 .These three systems (HIS, RIS, PACS) are interconnected and create the backbone for the radiology workfow in any RC.
All participating radiologists use a dictation system for report writing using speech-to-text.This often demands meticulous corrections, prompting manual adjustments due to limited accuracy, which "costs time" (P01).Besides, some radiologists copied the text from the dictation editor into Microsoft Word (P01, P03).The Word document was subsequently saved within RIS.Sometimes, radiologists wrote the report directly into their RIS (P04).The fnal report was uploaded into the PACS.
In our study, radiologists rely on traditional paper-based materials alongside their technical tools.Across all the RCs we observed, a signifcant amount of paper-based records was evident.Radiologists routinely refer to these materials, including PSA values, progress charts for PSA over time, medical histories, and more.Additionally, pre-printed prostate sector maps are indispensable tools used by radiologists to manually mark afected areas during the diagnosis process.We also noticed a large poster in the radiologist's ofce at RC01 detailing and visualizing the diagnosis scheme for mpM-RIs according to PI-RADS oriented on Blondin, Schimmöller and Quentin, [15], and P01 mentioned that he occasionally refers to this poster for guidance.
In Figure 3, we witness an example of the dynamic and multitasking nature of a radiologist's work during the prostate cancer diagnosis process in RC02.This image illustrates the intricate and demanding nature of the radiologist's workfow as they navigate between multiple digital and physical artifacts to provide precise prostate cancer diagnoses.

External Communication and Collaboration
Workload emerges as the most signifcant stress-inducing factor within the workplace, irrespective of the center's scale, as per the accounts of all radiologists.The ratio of radiologists to the number of patients constitutes the primary instigator of the considerable workload.Additionally, radiologists are also engaged in various administrative responsibilities, further burdening their workdays and intensifying pressure and stress.The information workfow adds further hurdles to the daily operations of the radiology department.P03 complained: Often the radiologists don't get enough information, so you can't make a clear diagnosis because preliminary information is missing [...], and that makes a lot of extra work that would actually be unnecessary.It would be more relieving if the information fow was better and pre-fndings [from other doctors] were more accessible so that a reasonable diagnosis could be made.
In RC03, we also observed instances of missing information, such as the absence of a specifc PSA value in the documentation.Consequently, P04 needed to personally inquire with the patient about their PSA value.
Physicians frequently engage in discussions with their peers, particularly for intricate or unusual cases, seeking additional opinions to enhance diagnostic accuracy.Some doctors convene in person to review cases, while others resort to phone consultations.However, feedback from patients or referring doctors, such as histologists, pathologists, urologists, and oncologists is seldom received by radiologists.For instance, in RC01, doctors mentioned that their hospital does not perform biopsies, leaving them without information about the accuracy of their diagnoses.Similarly, in RC02, while radiologists require histology/biopsy results, they rarely receive any feedback on them.Information is vital not only at the onset of diagnosis as input but also afterward as feedback or validation.
As P03 expresses, "It is annoying that we rarely get feedback from colleagues about the diagnosis, even by asking about it in the report, or via phone".P05 concurred, sharing that he also does not receive feedback from pathologists, even after requesting it.
In certain instances, when the radiologist requires information from the urologist, multiple inquiries are often necessary, demanding signifcant efort and consuming valuable time.Similarly, P04 expressed his concern by stating "radiologists specifcally ask about the result of the biopsy to evaluate their diagnosis, but urologists don't respond".Furthermore, P05 occasionally contacts referring physicians, and in approximately 20% of cases, he gains insights into the presence of carcinoma.However, he emphasized the need for consistent feedback across all cases.P05 indirectly gauged the quality of his diagnoses when he observed that many local urologists prefer referring patients to him over others.He mentioned, "That means that the failure rate can't be all that bad.But I don't have any objective data on that".During the interview, P05 continued to emphasize that incorporating feedback into the standard workfow is crucial.
Furthermore, P03 highlighted that after delivering his diagnosis, he often remains unaware of subsequent patient actions, whether they followed his recommendations and proceeded to a urologist for a biopsy or not.On occasion, returning patients arrive for further evaluation at a later date, becoming the only source of information about the patient's progress.P03 noted: "Since it is a lot of efort, unfortunately, a lot of data gets lost in the current process."

DESIGN IMPLICATIONS AND CONCEPTIONS OF AI
In this section, we delve into RQ2, "What potential does AI hold in enhancing radiologists' current practices for prostate cancer diagnosis?".Drawing on the insights garnered from our empirical research presented in section 4, we uncover specifc challenges that radiologists encounter in their current practices, resulting in resource bottlenecks.These bottlenecks pose signifcant obstacles to efciency and can disrupt the seamless diagnosing process.As we have learned from the literature (see section 2), AI might potentially enhance the diagnosis process in radiology.Hence, to tackle current challenges, we outline design implications and formulate AI concepts tailored to reshape future work practices.This section provides insights derived from related work as well as our observations and conversations with radiologists regarding their attitude and trust toward AI, and the potential integration of AI and automation to strengthen the efciency of the prostate diagnosis process.All the radiologists unanimously viewed AI as a promising avenue for enhancing their diagnosis capabilities as they envision AI systems ofering preliminary diagnoses and aiding throughout the diagnosis journey.This assistance spans a spectrum of functionalities, including employing visualizations to pinpoint and localize anomalies, automating complex calculations and reporting, ushering in a hybrid intelligence-driven approach, and facilitating multidisciplinary communication for prostate diagnosis.

Attitudes and Trust toward the Use of AI in Radiology
When queried about the satisfaction with current systems within the unique diagnostic frameworks of the RCs, all radiologists expressed contentment.Despite their overall satisfaction, they suggested that AI could provide additional support to further enhance their diagnostic processes.Each radiologist in our study was already familiar with and employed AI in radiology (for other cases than prostate cancer).P05 mentioned: The P04 mentions that AI is not intended to make a decision, but it can support and draw attention to certain areas where a lesion may have been detected.Therefore, P04 and P05 confrm a positive attitude toward AI but also state that humans should still be involved in the process as experience and consideration of the patients' context are required.For example, in the diagnosis, both volume and PSA density are considered.Even without a visible tumor on the MRI, values like pathological PSA progression, as mentioned by P04, can suggest the presence of a tumor.
Even though radiologists generally hold a positive attitude toward AI, they also recognize certain challenges.P05 acknowledged challenges in the implementation of AI in radiology (see section 6).P03 pointed out the complexity of diagnosing prostate cancer in a standardized manner, as suggested by the PI-RADS scheme.The diagnosis can vary among radiologists based on their individual experience and opinions, leading to diverse results.Teaching AI to learn beyond PI-RADS is challenging due to the non-uniformity of human evaluations used for the training data.
Research about the advantages of AI in healthcare is expanding the knowledge of doctors and boosting their trust and acceptance of AI systems.When asked if the radiologists would trust AI systems for diagnosis, all of them responded positively.However, the accuracy of AI systems and the corresponding trust of doctors are not straightforward.P03 had negative experiences with AI in brain MRIs, urging the importance of being able to validate AI accuracy.Nevertheless, he also knew from colleagues that AI can be very helpful in lung diagnosis as it hardly misses lesions in the lung.
The evaluation of detected lesions has a higher error rate, therefore he would trust the AI in detecting lesions, but not in interpreting these lesions as pathological or not.If the AI indicates no detection, P03 would trust the AI system.P01 and P02 also mentioned that they need to test the system frst for its accuracy in automation, and depending on the success of the result they would trust the system accordingly.

AI-Driven Visualizations for Prostate Diagnosis
Drawing from empirical research outcomes, we have conceptualized a utilization of AI in the realm of prostate cancer diagnosis.Our presented ideas are subsequently derived from the user needs and visions mentioned by the radiologists.Techniques centered around visualization hold signifcance in the context of prostate cancer diagnosis.Presently, we are actively engaged in the evaluation and implementation of such visualization approaches, while simultaneously conducting training for an AI prototype.Our innovation denoted as PAIRADS 5 strives to alleviate the responsibilities of radiologists through the integration of AI.This integration involves harnessing image recognition capacities in the interpretation of MRIs, serving the subsequent objectives: (1) Identifcation of prostate location, (2) Segmentation of distinct prostate regions, Detection and localization of potential lesions, (4) Classifcation of identifed lesions following the PI-RADS scheme.
To implement these ideas and features we are utilizing standard ML methods, as discussed in section 2. Through the collaboration with the radiologists, MRI images were provided to us, and for most of them, labels were added, from which the AI learns the recognition of the segments.We applied the Semi-Supervised Learning (SSL) approach to train our AI model considering SSL exhibits superior results by learning directly from a small set of labeled data and a large set of unlabeled data [74,115], and can lead to promising results in medical image segmentation [81].
As presented in section 2, CNNs excel in various domains, particularly in image processing tasks like image recognition, semantic segmentation, and object detection [7].Remarkable progress in medical image classifcation has led some CNN-based research to achieve performances on par with human experts [94].Leveraging the Deep Convolutional Neural Network (DCNN) for medical image processing is advantageous due to their prowess as feature extractors, eliminating the need for intricate and expensive feature engineering [65,118].Applying DCNNs as a semi-supervised approach to train AI models with prostate MRI images ofers distinct benefts considering these networks possess the capacity to grasp intricate features, reduce reliance on labeled data, demonstrate robust generalization capabilities, and adeptly capture both local nuances and global context [10,81].
Here it is crucial to highlight that the AI is designed to operate in a semi-autonomous capacity, providing support rather than completely taking control.Consequently, it assumes certain assistance tasks that can subsequently be passed to the radiologist for additional assessment and ultimate decisions such as the fnal decision on the severity and further treatment options.Therefore, segmentation and localization might be helpful and can also be done using techniques that come from the intersection of DL with object detection and semantic segmentation in computer vision, with post-processing to tell the objects apart.
P01 mentioned that "it would be nice if you could see the contours of the prostate with one click" as it takes around 10 minutes to do so by hand.Our results suggest that various visualization methods, such as outlining the prostate and highlighting specifc areas like the PZ or TZ through contouring or color-coding on the original MRIs, can enhance the readability and comprehension of the MRI.A (digital) prostate model is important to diferentiate between the zones of the prostate (P01).Figure 4 shows the current results of the AI prototype for prostate segmentation applied to original MRIs.The images are displayed in the LPS (Left, Posterior, Superior) orientation since most DICOM viewers use this orientation.As a frst concept, we can already see how the AI can distinguish and segment anatomical zones within the prostate.At this point, due to the relatively small amount of training data, it is not yet confdent that the AI can reliably segment all zones, but for the determination of the PZ, this has been successful so far.At this point, it should be noted that we will not go into full technical details within the scope of this paper (see section 6).
Visualizations could also be helpful in highlighting identifed lesions by outlining around the lesion's border as well as deciding which size and what type of lesion is present.Our observation in RC01 showed that the "clarity" of the margins of the lesion (e.g., a nodule) is crucial.If a margin is clearly visible, it is rather harmless as tumors tend to have blurred margins due to their scattering nature.This may be a design implication of the AI, so identifying and distinguishing abnormalities could be supported by visual highlighting (RC01).For the classifcation of the lesions, the size of the lesion is one of the factors used to determine the severity (RC01).This could also be automatically calculated by the AI and visually highlighted so that classifcation in the PI-RADS scheme (1-5) is suggested.Also, P05 agreed by saying: "Ideally, the system should provide you [..] both the detection and directly a classifcation".
Figure 4 shows a possible illustration for suggesting the PI-RADS score.This textual information on the scores for the lesions has been added manually so far.Accordingly, these classifcations are not yet implemented and cannot be defned automatically by the AI.Here, we argue, that the ofcial PI-RADS assessment approach can be used as an orientation for the training process of the AI [90,108].
P01 taught us that prostates vary in shape and size among individuals, making it challenging to rapidly discern the prostate through MRI scans.Thus, a system that can automatically detect the prostate's boundaries would signifcantly reduce time constraints and aid in accuracy (P01).While tumors are generally less common in the TZ, radiologists often face challenges in locating them (P03).
The evaluation of TZ is a difcult task for radiologists (P01, P05) as P05 emphasized "That's where AI will have to tackle it".In RC03, we observed that P04 could also not accurately identify whether an abnormality was in PZ or TZ.In a diferent more complex case Lesion Markings (red) According to [33,34].Segmentation Markings (yellow, blue) Trained According to [85] was a minor abnormality identifed, but it was unclear what it was.P04 would consult an external colleague because the lesion was in the TZ and therefore unlikely to be a tumor from his perspective.P05 presented challenges based on the example of an AI for lung cancer detection.Depending on the sensitivity setting of a threshold for detection, the AI either found too many or too few lung lesions.He claimed: "If you set the threshold too low, then it found no lesions, if the threshold was too high you suddenly found so many lesions that it then doesn't help you anymore."Therefore, he viewed the challenge of adapting algorithms accordingly to strike a balance that ultimately transforms them into efective support.
Additionally, it can be advantageous to visualize prostate MRIs in a 3D format, as this approach provides radiologists with improved insights into both organ structures and potential abnormalities.In instances where image orientation requires verifcation and adjustment, radiologists commonly resort to a manual process using a ruler tool.However, as P04 emphasized, the integration of a 3D feature has the potential to address this challenge.
As suggested in section 2, the decision of the AI system will be made more understandable and transparent through visualizations, which will lead to a fair and responsible perception of the humansystem decisions [72].This can lead us forward in a discussion of whether visualization can be the frst step of explainability.It was found that explanations were sometimes seen not as directly aiding physicians' AI comprehension, but as flters or tools to streamline understanding and reduce interaction time [116].

Use of Automation for Manual and Redundant Tasks
The current practice unveils a multitude of manual tasks and redundancies.Our observations indicate that the calculation of prostate volume, PSA density, and lesion diameter, the reporting, and the entire data transfer process throughout the diagnosis by radiologists have long been plagued by manual and redundant tasks, making them both time-consuming and error-prone.The introduction of an automation tool for executing these tasks would yield signifcant advantages.Below, we outline the challenges radiologists encounter while performing these tasks and how AI can help address these challenges.

Manual Calculations.
As discussed previously, a notably signifcant manual task involves the calculation of prostate volume and PSA density.An automated calculation can be benefcial in this context.Every radiologist recognized the beneft of automated volume calculation or even more, such as P03 emphasized a great advantage in the use of AI for tasks that can be automated such as segmentation calculation of volumetry, PSA density, etc.Given the current practice with radiologists manually measuring the three sides of the prostate using a distance meter and inputting these measurements into a calculator for volume calculation, we propose that an AI system could signifcantly assist by automatically identifying and localizing the prostate through image recognition.In agreement with P03, an automatic and precisely identifcation of the prostate border and calculation of the volume as well as the PSA density (if the PSA value is provided), which might be transferred into the report would alleviate the burden of manual labor and greatly streamline the reporting process.However, it is crucial that this automated process does not take longer than what radiologists currently spend on the task to maintain efciency.

Reporting.
During the reporting phase, our observations revealed the presence of redundant tasks that impact efciency.As discussed in section 4, radiologists adopt varying methods to draft reports, such as repurposing previous patient templates, necessitating manual updates, or employing multiple text editors.Additionally, the use of voice dictation to transcribe critical measurements and calculations is commonplace, yet time-consuming.These manual processes consume precious moments that can be avoided through automated structured reporting (P01, P02).As described in section 4.1.4,a prostate diagnosis report includes a prostate sector map, in which the identifed lesions are currently drawn manually.P01 argued "It would be useful if one could digitally draw the regions directly in Word [diagnosis report]".P04 also meant that "It would be great if this could be automated in the report." Hence, there seems to be an urgent user need for automating lesion drawing in the prostate scheme (P01, P03).P01 and P02 concurred with the need for a drawing tool to mark areas of interest directly within the prostate model and the transfer of these markings into the report, potentially via a form or prostate diagram that can be flled in.As a possible solution, Figure 5 shows how in our AI prototype the identifed lesions (marked red) are automatically transferred to the PI-RADS sector map so that no manual transfer is necessary.[33,34] In our observations, it was evident that each radiologist adopted a unique reporting method due to the absence of a standardized system.However, all radiologists emphasized the necessity of a uniform reporting framework.The establishment of a standardized reporting system for prostate cancer diagnosis demands thoughtful planning to cater to radiologists' needs and address their challenges.The system can feature a comprehensive customizable template containing patient information, history, imaging details, fndings, interpretations, and recommendations.The inclusion of pre-defned text modules in the template could facilitate efcient reporting.A robust speech-to-text can also be integrated for data entry by accurately transcribing necessary spoken input.Integrating imaging data directly into the reporting platform can enable seamless reference to images, eliminating the need to switch between applications.Therefore, the reporting system can be embedded with the existing essential systems, including HIS, RIS, and PACS.Taking various measurements and performing related calculations, as discussed in section 5.3.1, can be automated with the results seamlessly integrated into the report without manual intervention, thereby improving overall efciency.The system should incorporate PI-RADS to ensure uniformity in conveying fndings.The sector map should be incorporated into the reporting system as well so that it could serve as a reference for indicating the exact location of lesions or suspicious areas detected in the MRI images.The inclusion of such an automated standardized reporting system has the potential to save radiologists time, enabling greater focus on diagnosis and patient care.

Internal Data Transfer.
In terms of data transfer, the MTRAs need to manually upload the MRIs into the archive system, PACS (RC01-RC03).The accuracy of the data transfer needs to be checked again manually by the radiologists since a mistake can have critical consequences.MRIs will be automatically deleted if they are not archived in the PACS.Also, the MRIs will be copied on CD for the patients from the PACS, and the report is sent to the administration via the PACS.The administration needs to scan paper-based artifacts (e.g., the prostate sector map with manual markings of the afected area) for archiving and send the report (a letter) and additional artifacts (prostate model) via post to the responsible doctor.These redundant work steps and manual tasks consume signifcant time and efort, leading to dependency on radiologists for certain steps that need optimization.With AI-driven data extraction and dissemination, pertinent patient information can be automatically extracted from records and distributed across necessary channels, eliminating the need for repetitive data entry.Utilizing AI to optimize internal data transfer not only alleviates radiologists' administrative responsibilities but also enhances data accessibility, streamlines workfows, and ensures relevant information availability across various platforms for diagnosis and decision-making.

Amplifying Human Capabilities through Hybrid Intelligence
By analyzing the current context of use, we focus on HCAI design to research human variables and uncover medical acceptability hurdles to promote a transformational human-AI collaborative relationship as a way of hybrid intelligence [62] centered on prostate cancer diagnosis.We observed variations in the radiologists' artifacts and workfows across diferent RCs, with a key distinction being whether the diagnosis was conducted individually or through a double diagnosis approach.In the majority of cases, the prostate diagnosis is conducted by a single radiologist due to resource limitations.Many RCs have only one experienced radiologist capable of diagnosing prostate MRIs.Even if additional radiologists are available, they often lack experience in prostate MRIs and thus cannot ofer a second opinion as we observed in RC02 and RC03.
Our fndings indicate that while double analysis can ofer advantages in theory, its practical implementation is challenging due to resource constraints, including time limitations and a scarcity of available radiologists which we mentioned in the introduction.AI holds promise in augmenting cancer diagnosis for radiologists by enhancing the precision, efciency, and uniformity of prostate cancer detection.As such, we propose that the secondary evaluator need not exclusively be human; AI can assume the role of an assistant, generating preliminary or supplementary fndings.
We should not replace human performance but enrich their ability to diagnose by providing them with a second opinion through AI as a concept of hybrid intelligence.Ultimately, the decision always lies with the human entity, the radiologist.Nonetheless, an AI can certainly serve as a technical assistant that can ofer an additional perspective and pinpoint potential lesions, efectively providing a second opinion.This aligns with the fndings of the qualitative study by Griesshaber et al., which explains the signifcance of the AI system being adequately capable of acting as a supportive tool to direct users toward specifc (suspicious) areas [47].
In our research, we specifcally inquired about the preferred order of the diagnosis, whether the AI should provide a result before the radiologist examines the MRIs or after.Our fndings reveal varying opinions regarding the sequence of the diagnosis process.Some believe that the order does not signifcantly impact the outcome, while others assert that the AI's involvement should come after the radiologist's evaluation.P03 emphasized that the ultimate diagnosis will be made by the radiologist regardless of the order.P05 advocated for the radiologist to diagnose frst and then consult the AI.He drew from the example of spinal diagnosis, where radiologists initially assess the images without AI infuence, followed by incorporating the AI's result.Consequently, the reporting process commences with human evaluation, followed by an AI-assisted evaluation supplement.P05 found this approach useful, expressing concern that the reverse sequence might lead to a less meticulous review when radiologists realize the AI is at play after several diagnoses.He also emphasized that AI is not meant to diagnose alone.

Multidisciplinary Communication and Decision Support Tool
As discussed in section 4.5, there is a signifcant lack of communication among doctors.Conversations with radiologists highlighted a crucial issue: the fragmentation of medical disciplines, encompassing radiology, urology, and pathology.The prevailing communication challenges and inadequate information exchange across these specialties can result in misinterpretation of observations, ultimately causing diagnosis delays and potentially erroneous treatment choices.By soliciting feedback on specifc user requirements to enhance radiologists' workfows, every radiologist in our study expressed a desire for a multidisciplinary system.For instance, P04 envisions an automatic feedback system, a cloud solution where urologists, radiologists, and pathologists can access each other's patient data online, facilitating collaborative decisions like whether to perform a biopsy based on factors such as PSA values or not.
P05 underscored the signifcance of establishing a connection between pathology and radiology.In instances where patients have undergone radical prostatectomy, he emphasized the potential for substantial quality enhancement by conducting histological mapping of the prostate and aligning it with image data.This, in turn, addresses the reliance on subjective decisions that lack objectivity.Hence, he suggested collaborating with the pathology department that evaluates prostates.An integration of pathology results and MRI images provides a more holistic understanding of the patient's condition.Refecting on the topic of collaborating with other doctors, P04 agreed by saying: "The only thing that matters is what everyone decides together.That's the foundation." Based on our empirical fndings, the current practice involves sending out reports and images to external doctors via postal mail, often including a CD.P05 was the exception, as he employs a digital system to transmit images as well.P03 expressed a desire for a streamlined process, envisioning the convenience of sending everything with just a single click, either through email or a multidisciplinary system.Similarly, P04 emphasized the benefts of digitization and advocated for transitioning to an online platform for communication.In general, every radiologist we interviewed discussed the implementation of a multidisciplinary system involving various medical professionals and patients positively.
In 2019, Heidenreich et al., [52] highlighted in their study the advantages of a multidisciplinary approach, where diverse experts contribute their insights to collectively determine the optimal treatment strategy for patients.Schlemmer [98] says that the cooperation between urologists and radiologists is important to control and improve quality in the long term.Informed and collaborative decision-making involving radiologists, urologists, pathologists, and other specialists is pivotal in shaping personalized treatment plans for prostate cancer patients.Thus, enhancing prostate cancer diagnosis requires the implementation of improved communication and collaboration among these disciplines, ultimately enhancing diagnostic accuracy and treatment quality.An overarching communication platform, such as a joint tumor board, could be established, integrating clinical data, imaging, and pathology results.This platform would facilitate specialists in sharing expertise, collaborating on cases, and refning treatment decisions to enhance individualized patient care.

DISCUSSION
The use of AI clinical decision support systems in various forms of medical diagnosis has been of increasing interest [41,54,69] and indicates that this research interest and the relevance of our fndings, is not exclusive to Germany or radiology or prostate cancer.In particular, qualitative studies, like ours, point to a range of diferent attitudes and concerns amongst clinicians [3,57,58] such as lack of evidence of efcacy, and lack of explainability.As King et al. write, "AI is more likely to be accepted if pathologists are able to 'make sense' of the technology, engaged in the adoption process, supported in adapting their work processes, and can identify potential benefts to its introduction" [69].Despite AI demonstrating great promise in healthcare and medicine, much of this development remains unimplemented due to ML models being trained on quasi-data and assessed only in controlled experiments that differ signifcantly from real-world implementation circumstances [87].Therefore, understanding the actual context of use, existing infrastructure, utilized artifacts, and systems, including their current challenges, becomes essential.Through our empirical work following a practice-centered design approach, we gained insights into current work practices, where we identifed numerous crucial aspects of radiology for prostate cancer diagnosis.
The process of diagnosing prostate cancer involves various manual tasks and is error-prone requiring additional time and efort for correction.Additionally, extended day shifts, substantial workloads, and the complexity of the tasks can contribute to mistakes.We have also observed signifcant manual involvement in tasks such as calculations during diagnosis, reporting, and information transfer across diferent systems, resulting in errors made by the involved users.To avoid human errors, AI can be used to be more efcient and precise.AI has the potential to impact many aspects of radiology, with a special emphasis on oncology applications, including utilizing image-processing technologies to improve diagnostic accuracy and efciency in the feld [1,54].
During our feldwork, we encountered a noticeable disparity between the recommendations put forth by various research projects and guidelines regarding the diagnostic procedure for prostate cancer, and the actual practices within radiology.According to the current PI-RADS guidelines, the diagnostic process should be consensus-based, involving measurements by two independent observers [108].However, due to the shortage of radiologists and the increasing patient workload, only two RCs (RC01 and RC04 mentioned by P05) implement the double analysis process, while in other centers, a single radiologist performs the analysis alone.
While the PI-RADS classifcation serves as a guideline to aid radiologists in diagnosing prostate cancer, it is important to recognize that its theoretical clarity does not always translate to practical assessments.The complexity of prostate cases defes simple categorization, often existing in nuanced shades of gray rather than stark black-and-white distinctions.The process of prostate cancer diagnosis guided by PI-RADS is intricate and time-intensive.Other critical factors, including the radiologists' extensive experience, refned evaluation techniques, and high-resolution image analysis, come into play when diagnosing prostate MRIs through PI-RADS.Radiologists' individual expertise contributes to variations in diagnosis, infuenced by considerations such as patient history, pre-diagnosis, and age.The intricate interplay of these multifaceted factors highlights the complicated nature of prostate cancer diagnosis and calls attention to the need for a comprehensive approach that goes beyond rigid classifcations.
Currently, there is no progress tracker for the radiologists to know about the current stage of the patient and also to verify their diagnosis.We have mentioned the current situation regarding communication among diferent doctors in section 4.5.Cooperation among doctors in cancer is unavoidable.Considering the benefts for patient survival, especially among older patients, a Multidisciplinary Team (MDT) has been recommended as the standard treatment approach in modern oncology [92].The MDT comprises specialists from various felds who convene regularly to review patients' diagnoses and treatment plans.This approach is considered as 'best practice' in cancer care, both from a legal and ethical standpoint [28].The German Guideline program in oncology also recognizes that determining the medical treatment strategy for prostate cancer should involve an interdisciplinary and multiprofessional approach [73].However, in medical practice, the MDT is not as well-established as German national guidelines recommend [79].
Throughout our empirical work, we observed varying perceptions of AI between researchers and users.Our study initially aimed to uncover the potential of AI in prostate cancer diagnostics through an analysis of current work practices and discussions with radiologists.Interestingly, as we delved into specifc user needs and requirements, we realized that many ideas for assisting radiologists with technology did not necessarily require AI but rather 'simple' automation (see section 5.3).This divergence in the understanding of AI appears to exist among diferent stakeholders, including researchers, developers, and radiologists, as noted by other researchers [100].This shows the relevance of human-centered design, as we only identifed the diferent perceptions by working closely with users.
The collaborative engagement with radiologists has revealed their overall receptiveness toward AI support.It is emphasized that AI cannot supplant the human factor, particularly the wealth of radiological expertise.The complete emulation of current practices with AI can prove challenging, as decisions often stem from experience and lack well-defned, concrete criteria.For an experienced radiologist, discerning a tumor can at times rely on intuition, defying objective quantifcation.Occasionally, it hinges on subtle details, such as the holistic image impression or borderline cases, which prompt radiologists to make unconventional judgments.Compiling these criteria into an algorithm that attains a level of validity robust enough for 100% reliance poses a formidable challenge, as perceived from the radiologist's standpoint.Radiologists possess a specialized skill known as 'professional vision' [45], which involves their ability to diferentiate between normal and abnormal fndings by recognizing typical visual patterns and employing specifc techniques to highlight discrepancies [93].Hence, ultimately, the AI should serve as an assistant, fostering seamless collaboration with the radiologist to achieve the most comprehensive diagnosis possible through the fusion of humans and AI.
Our research also highlights several issues connected to accountability and trust, especially concerning the design, redesign, deployment, and evaluation of technologies [93].As new tools integrate into complex organizational workfows, we must address the challenges of their integration and adaptation to existing work practices.Since achieving trust from physicians is important for the adoption of AI [110], we need to consider particular design features of diagnostic tools that might either instill 'trust' or provoke 'mistrust' among clinical professionals.

Limitations and Strengths
While our study ofers valuable insights into the perspectives and practices of radiologists regarding prostate cancer diagnosis, certain limitations should be acknowledged.Firstly, our sample size was fairly small, which may restrict the generalization of our discoveries to a larger radiologist population.Furthermore, we might not have been able to completely represent the specifc contexts and characteristics of the RCs we were engaged with and may not fully capture the diversity of practices across various healthcare settings.Though the sample size was small and lacked diversity, we could delve deeply into the nuances of each local RC which gave us comprehensive contextual insights.
Another limitation lies in the fact that the conceptions we have suggested remain at the level of ideas and have not been fully implemented or thoroughly evaluated.While these ideas hold promise for enhancing the diagnosis process and supporting radiologists, it is important to recognize that our proposed solutions are still in the conceptual stage, and their true value and feasibility can only be determined through practical implementation and comprehensive evaluation.Our future work involves refning these ideas, conducting thorough assessments, and addressing any challenges that may arise during implementation.
Additionally, the qualitative nature of our study opens it to subjectivity in interpreting the data.Since medical contexts can be complex and nuanced, researchers who are not well-versed in medical terminology or practices may unintentionally misinterpret certain aspects of the observed behaviors or discussions.To mitigate this limitation, we tried to collaborate closely with radiologists who provided insights and clarifcations, ensuring a more accurate interpretation of the collected data and allowing us to understand their obstacles and needs.Hence, our human-centered approach has provided us with valuable insights into the current practices and challenges faced by radiologists in the context of prostate cancer diagnosis.This practice-centered design lenses gave us ideas for supporting their current practice with the intent to aid radiologists in what they are doing best [95].Additionally, our use of contextual inquiries was invaluable, enabling us to systematically observe and understand the workfow in its discrete stages.As a result, the data were systematically analyzed, and through the data excerpts, we were able to give voice to our participants.
The interdisciplinary collaboration in our research team together with the local radiologists and AI developers helped to leverage diferent expertise and viewpoints.This collective efort has been instrumental in conceptualizing optimal solutions, as Greenhalgh stated: "It is not individual factors that make or break a technology implementation efort but the dynamic interaction between them" [46].This collaborative, multidisciplinary approach aligns with the perspective put forth by Kessler et al., who argue: "We need studies that are interdisciplinary, nondeterministic, locally situated, and designed to examine the recursive relationship between human action and the wider organizational and system context" [66].

CONCLUSION
In this paper, we provide insights from an empirical study with an HCAI approach emphasizing practice-centered design through which we gained knowledge about current practices and AI potentials in radiology within the context of prostate cancer diagnosis.
To address RQ1, we immersed ourselves in the feld and conducted contextual inquiries and interviews.These methods allowed us to acquire profound insights into the prevailing practices and challenges faced by radiologists in prostate cancer diagnosis.These insights aford us a comprehensive overview of the diagnostic process, which we delineated into fve primary phases: patient inquiry, image acquisition, image interpretation, reporting, and verifcation.We also garnered knowledge concerning the utilization of various artifacts and systems and internal and external communication dynamics.
Subsequently, to respond to RQ2, we extrapolated specifc user needs and design implications.We recognized that radiologists perceive AI as a promising avenue for amplifying their diagnostic capabilities.Hence, we conceptualized a prospective HCAI system to address the identifed challenges with features, such as identifcation of the prostate location, segmentation of distinct prostate regions, detection and localization of potential lesions, and classifcation of identifed lesions following the PI-RADS scheme.
We understood that AI does not represent a separate authority making decisions but should be used as an assistant to provide a second opinion to the radiologist, who often works alone, and thus can contribute to more accuracy and efciency.In our study, radiologists often emphasized there is no substitute for human agency because the radiologist's experience and expertise are hugely important in ultimately making a decision.Nevertheless, all participating radiologists express a positive attitude toward AI and are eager to collaborate with it to mitigate their challenges and reduce their workload.This includes using visualizations to locate lesions, automate complex calculations and reports, adopt a hybrid intelligencedriven approach, and facilitate multidisciplinary communication for prostate diagnosis.
Taking account of our empirical fndings, we conceptualized a system combining advanced features employing AI and image analysis techniques, with some straightforward automation functions.This AI-driven system can assist radiologists in analyzing mpMRI images for diagnosing prostate cancer with the help of visualizations, and ultimately make the fnal decision that ensures human intelligence rather than replacing it, resulting in hybrid intelligence [62].This can help radiologists make more efcient, accurate, and timely diagnoses, which can ultimately reduce the radiologists' workload and improve patient outcomes.
For our future work, we intend to implement the AI, co-designed by radiologists and thoroughly evaluate its real-world efectiveness in addressing challenges during prostate cancer diagnosis.Therefore, through appropriation studies, we will investigate integrating our solution into existing healthcare infrastructures, where the evaluation will involve working closely with radiologists to assess the system's capability, accuracy, and explainability.
Heinsberg, Radiologie Rhein-Nahe, and Radiologie Ibbenbüren for their cooperation in our feldwork.Additionally, we thank our student assistant, Jessica Fraas for her support in data transcription.

Figure 1 :
Figure 1: Research process model of our empirical study

Figure 2 :
Figure 2: Workfow of a radiologist in prostate cancer diagnosis (for an enlarged version, please refer to Figure 6)

4. 1 . 2
Image Acquisition.MRI emerges as one of the most potent imaging modalities in contemporary clinical practice, carrying distinct clinical signifcance within radiology [109].According to the German guideline program in oncology

Figure 3 :
Figure 3: Radiologist juggling among diferent artifacts during diagnosis: Reviewing MRI images (left monitor), generating report (right monitor) using his voice dictator (on the table), and annotating on the prostate sector map (on the table)

Figure 5 :
Figure 5: AI result of localization of a lesion in the PI-RADS prostate sector map based on Turkbey and Rosenkrantz et al. [108].Image modifed through red markings based on Cuocolo et al. (2021)[33,34]

Figure 6 :
Figure 6: Enlarged View -Workfow of a radiologist in prostate cancer diagnosis PZ The Peripheral Zone (PZ) refers to the outer region of the prostate gland and is the most common location for prostate cancer to develop.3, 6, 10 Radiomics It involves using advanced computational techniques to extract and analyze quantitative data from medical images (like CT scans or MRIs).This method aims to fnd patterns and biomarkers within these images to aid diagnosis, treatment response prediction, and disease characterization.3 RC Radiology Center(s).2, 4, 5, 7, 9, 11-14 RIS Radiology Information System.7, 11 RML Radiomic Machine Learning.3 RQ Research Question(s).2, 4, 8, 14 SSL Semi-Supervised Learning.9 T1W A T1-Weighted (T1W) image is a type of MRI sequence that emphasizes diferences in the spin-lattice relaxation time (T1) of tissues.It provides detailed anatomical information by highlighting variations in tissue contrast, particularly useful in visualizing diferent soft tissues within the body.5, 6 T2W A T2-Weighted (T2W) image is a type of MRI sequence that emphasizes diferences in the spin-spin relaxation time (T2) of tissues.It enhances contrast based on tissue water content and is valuable in imaging anatomical structures, especially in visualizing fuid-flled spaces and abnormalities within tissues.3, 5, 6 TZ The Transitional Zone (TZ) refers to a region within the prostate gland that surrounds the urethra.3, 6, 10 XAI eXplainable Artifcial Intelligence.3