Designing a human-centered AI solution with radiologists to improve prostate cancer diagnosis

2.1 Context study

We conducted qualitative research by working with different radiology centers across Germany for 2.5 years. We had 11 contextual inquiry sessions, ⁴⁰ involving seven radiologists and other medical staff members, each lasting between three and five hours. Furthermore, we conducted 11 semi-structured interviews with radiologists, with a duration ranging from 40 to 80 minutes. Additionally, two workshops were organized: first, an ethics workshop with seven participants (excluding the first two authors) from the fields of ethics, radiology, HCI, and AI development, and second, an international workshop on Human-Centered AI in Healthcare held at the ECSCW conference in 2022, which involved 10 participants (excluding the first three authors) with backgrounds in HCI/CSCW, healthcare, and AI development. ⁴¹

Through the context study, we gathered in-depth insights into the daily practices of radiologists. ⁴² We learned about individual steps within their workflows, contextual factors influencing the practice, usage of various artifacts and systems, internal and external communications, practitioners’ needs, practitioners’ challenges, and their perspectives towards AI as a potential assistant. This in-depth analysis aimed to establish a solid foundation of understanding and to define requirements, serving as the basis of the development of our HCAI solution. Part of our insights has been published in the proceedings of the CHI conference in 2024, ²⁸ where we discussed a section of our iterative empirical study on radiologist practices in diagnosing prostate cancer and how AI can enhance decision-making in that context. While additional empirical insights emerged through our iterative work, which are not yet published, we will not discuss them in detail in this paper, as it focuses on the design study. But to comprehend our design decisions, we will briefly present key insights from the whole context study, including an overview of current prostate cancer diagnosis procedures using MRI, the user needs, and requirements for an AI tool to support radiologists in the diagnostic process.

2.2 Requirements of the AI-tool

The insights and challenges we have identified through our context study are transformed into user needs and requirements. According to the often-mentioned user needs, we have prioritized the most relevant requirements:

1. Identification of the prostate location

2. Segmentation of distinct prostate regions

3. Detection and localization of potential lesions

4. Classification of identified lesions following the PI-RADS scheme

5. Calculation of prostate volume and PSA density

6. Standardization of the reporting system

7. Giving feedback about AI results

Later, in explaining the prototyping process in Section 5, we will refer to the requirements numeration, which doesn’t imply a priority.

3.1 AI and decision-support systems in radiology and prostate cancer diagnosis

AI is being used more and more for a variety of activities, but decision support is the primary application it serves. ⁴⁷ AI-driven decision aids are increasingly adopted in hospitals and clinics to support medical practitioners in analyzing medical data and interpreting results more accurately. ⁴⁸ Clinical Decision Support (CDS)/Clinical Decision Support Systems (CDSS) are developed to offer timely, evidence-based recommendations for diagnosis and treatment. ^{49 , 50 , 51} With advanced AI AI-based Medical Diagnosis Support Systems (AIMDSS) are growing into a significant component in pathology and radiology, ⁵² while Computer-Aided Detection (CAD)/Computer-Aided Intelligent Diagnosis (CAID) shifted from just detecting suspicious regions to also interpreting and diagnosing those regions. ^{49 , 53 , 54}

AI has the potential to revolutionize radiology practice by offering valuable support to radiologists in various aspects and steps of their workflow. ^{25 , 55 , 56} There exists a variety of studies on the use of AI for prostate cancer diagnosis. As the prostate exhibits high variability in shape and appearance, Convolutional Neural Network (CNN) can better cope with these issues, gaining popularity for segmentation. ⁵⁷ Many examples ^{58 , 59 , 60 , 61 , 62 , 63 , 64 , 65} demonstrate CNN applications in medical imaging for prostate cancer detection and segmentation, with a primary focus on utilizing MRI images for image classification, analysis, and segmentation.

Litjens et al., developed an automated CAD for prostate cancer using MRI, comprising steps such as prostate segmentation, feature extraction, and candidate classification. Their evaluation showed superior performance, suggesting potential usefulness for radiologists in both first- and second-reader settings. ⁶⁶ To understand physicians’ views on AI, Buck et al. ⁶⁷ interviewed German physicians and medical professionals closely tied to radiology. They discovered that while specialists acknowledged AI risks, they tended to double-check its recommendations, leading to extra effort. One participant expressed frustration when AI missed details, requiring thorough reviews and negating potential time-saving benefits. Penzkofer et al., aimed to identify prerequisites for the successful implementation of clinically relevant AI in prostate MRI diagnosis. ⁶⁸ They noted limited adoption of AI in clinical practice despite numerous studies exploring its diagnostic potential. Regarding usability, they suggested tailoring user interfaces to radiologists’ workflows for comprehensive MRI evaluation alongside AI results.

Studies on the use of AI for prostate cancer diagnostics primarily focus on the technical perspective, describing the development of AI algorithms. The HCI perspective, which involves understanding user needs and desired interactions with the AI, is often overlooked. Our work addresses this gap by emphasizing strong collaboration with radiologists.

3.2 Human-centered design in healthcare

The integration of AI algorithms has encountered challenges in real-world applications when not considering insights and recommendations arising from the field of HCI. ⁴⁹ Users who lack an understanding of the system’s functionality and capabilities become overwhelmed by the system’s output, have limited situational awareness, and may experience a loss of control over the system. ⁶⁹ Procter et al., also highlighted that AI systems in healthcare often lack an understanding of the organizational context, limiting practitioner trust in their recommendations. ⁷⁰ However, integrating an HCAI assistant into the radiology workflow enhances result precision, visually simplifies analysis, and significantly reduces assessment time, thus boosting workflow efficiency for radiologists. ⁷¹

Following an HCD process, the current context of use can be fully analyzed to define requirements for the new AI tool, ensuring effective, efficient, and satisfactory user utilization. Particularly in the field of healthcare, where sensitive data is processed, considering human factors and putting humans at the center of an AI design process is the key to good usability and users’ acceptance of the system, ⁷² especially when it comes to designing an eXplainable AI (XAI), ^{47 , 73} emphasizing the need for interfaces to offer algorithmic decision descriptions, multi-layer rationalization, and data origin information. ^{74 , 75} This supports increased transparency and trust between users and the AI system, which are found to be decisive in the adoption of such systems. ⁷⁶ The successful deployment of AI technologies hinges on their seamless integration into existing clinical workflows and infrastructure, preserving tasks and processes unaltered. ^{77 , 78 , 79} Understanding the current context of use is crucial for ensuring positive adoption. ^{77 , 80}

Socio-technical, participatory methods have been recommended by an increasing amount of research to effectively involve domain experts throughout the AI development lifecycle. ^{81 , 82 , 83} Ooge et al. ⁸⁴ emphasize the vital role of domain experts and end-users in developing visual analytics tools, aligning with recommendations from HCI scholars. Health professionals’ feedback underscores the efficacy of participatory and user-centered interaction design methods, particularly in UI drawings, facilitating requirement description and common understanding with system developers. ⁸⁵ Increased user involvement in design is crucial, with user input identified as pivotal in the success or failure of complex technology. ⁸⁶ 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, ^{87 , 88} can contribute toward more ethical, adaptable, and useful AI systems. ⁸⁹

Usability is crucial for integrating medical software into workflow and adoption. ⁹⁰ HCD methodologies facilitate achieving usability in medical technology. ⁹¹ DT, known for its empathy, creativity, and collaboration, is effective across healthcare domains, offering user-centered solutions. ³¹ Its success in healthcare stems from considering contextual factors, including user needs and clinical evidence. ⁹² DT outcomes in healthcare outperform traditional interventions in terms of usability and effectiveness. ⁹³ Given the importance of usefulness and ease of use in medical technology adoption, ⁹⁴ DT presents a promising approach. Additionally, Chen et al., emphasize a human-centered approach in DT for explainable medical imaging systems. ³⁰

4.1 Design thinking workshops and low-fidelity prototypes

We organized two dynamic and collaborative DT workshops to leverage interdisciplinary collaboration, fostering creative problem-solving and stimulating out-of-the-box thinking, particularly beneficial due to the time constraints faced by the radiologists and their widespread distribution. Since there are multiple models available illustrating the process of DT, ^{95 , 96} we have adapted the DT process according to our available resources, such as time, place, and participants (see Subsection 7.1). Our goal was to generate innovative ideas for an AI-based diagnostic tool tailored for prostate cancer diagnosis and explore visual concepts for different features.

The first workshop was conducted online to integrate participants from different locations. The session lasted for 150 min, bringing together a diverse group of 11 participants considering innovations in healthcare come from diverse user types. ³¹ The participants included three radiologists experienced in prostate cancer diagnosis, three developers with AI experience, two UX professionals, and three HCI researchers (including one HCI research assistant). For recruiting, we used social media platforms such as LinkedIn^[2] and Meetup^[3] to advertise the workshop as an event and to explicitly motivate AI developers, radiologists, UX designers, and HCI enthusiasts to participate. While social media effectively reached most roles, radiologists were specifically invited through direct contact. Some radiologists expressed interest in our research project after learning about it through the German Radiological Society, allowing us to extend explicit invitations to the workshop. Most of the participants did not know each other before the workshop. We used Webex^[4] to meet online and Miro^[5] to gather our insights. We divided the participants according to their roles into two multidisciplinary break-out rooms to work on the tasks together and facilitate diverse discussions.

We structured our workshop into five phases, namely Empathize, Define, Ideation, Prototyping, and Evaluation, ⁹⁵ to collaboratively engage with our participants. First, we explained the context of use and described the key findings from our context study, which set the focus for the workshop. Then we requested insights from the radiologists in each group to discuss challenges and limitations in their current practices regarding our focus and share their expectations about an AI-driven tool with the group members, so that the participants can gain an empathetic understanding of the user needs. Our previously identified user needs and requirements were confirmed by the radiologists through these steps in the workshop.

We delved into understanding the users’ perspectives using an empathy map. Building on that, we crafted actionable problem statements as Point-of-View (POV) statements to help us define specific needs and core challenges of our users. Afterward, we defined How-Might-We (HMW) statements as open-ended questions to generate multiple ideas focusing on desired outcomes. Then we created sketches for rapid ideation by identifying innovative solutions to the problem statements. After finalizing the best parts of the sketches using Dot Voting, we focused on prototyping by producing low-fidelity versions of the tool. Participants were instructed to visually map out the interface or key functionalities of the prototype based on the MRIs we provided using the Wireframe tool on Miro. We evaluated the prototypes made by each group with a feedback session with the whole group. Though we discussed together after each of the phases, the final discussion ensured a comprehensive exploration of ideas and the refinement of solutions.

The second workshop was done in person and spanned 90 min, bringing together 20 master’s students of HCI from our university. This approach aimed to leverage HCI and design perspectives during the creative phase, potentially fostering even more innovative ideas. The students were all part of an HCI course, so they knew each other. The insights gained from the context study and the feedback received during the first workshop served as the foundation for developing further design ideas. We didn’t show them concrete design ideas from the first workshop to avoid bias. We divided the participants into four groups consisting of five students, where at least one of them in each group has professional experience in UX design. Owing to scheduling restrictions, we scaled down the scope of our in-person workshop and concentrated mostly on the ideation and prototyping stages with the participants. An information document detailing the scope of our study was sent to them in advance, and to build empathy with stakeholders, we reaffirmed the context of use and the main findings at the beginning of the workshop. We provided four POVs, which were generated in the previous workshop, to streamline the ideation process, ensuring a focused exploration of specific problem areas within the limited time frame. Each group was given one unique POV, which was used by the participants to create multiple HMWs.

Following that, each participant brainstormed solutions for the HMWs they had chosen, utilizing the Crazy 8 method. Based on the dot voting on the best design parts, participants advanced to the prototyping step by using the provided materials to build four paper prototypes, one for each group.

4.2 Hi-Fi prototypes and evaluation

The six prototypes created by the participants at the end of the two DT workshops, two from the first and four from the second, were not intended as finished solutions; instead, they served as a foundation that our research team, involving HCI researchers and AI developers, further merged and refined through brainstorming sessions while considering participant feedback, translating the design ideas into a clickable Low-Fidelity (Lo-Fi) prototype on Figma.^[6]

Derived from that, we developed a Hi-Fi prototype, which was evaluated with five radiologists. Afterwards, we returned to a Mid-Fi Figma prototype to focus on iterative refinement of core features of the solution based on the initial feedback without being constrained by implementation details. This allowed us to rapidly integrate insights from the first evaluation and test design alternatives more flexibly. The Mid-Fi prototype was also evaluated with another five radiologists.

We have conducted initial evaluations through usability tests as a pre-stage to the appropriation study. The evaluation aimed to assess the usability of the AI-based solution by examining user needs and identifying potential flaws. Eight of the usability tests were conducted individually via Zoom,^[7] allowing them to participate remotely from their work environment. Three participants (P05, P06, P07) remotely completed approximately 90-min sessions each, accessing the Hi-Fi prototype hosted on a university server through a standard internet browser, along with anonymized MRI images for importing into their Digital Imaging and Communications in Medicine (DICOM) viewer. One 30-min session with two participants (P02, P41) was conducted on-site at a radiology center, where the radiologists accessed the same prototype directly on the researcher’s laptop. Additionally, five sessions of approximately 45 min were conducted remotely, where participants (P04, P06, P07, P08, P09) accessed the Mid-Fi prototype through a shared Figma project.

While the remote participants shared their screens and live video during their exploration, participants on-site were directly observed. All participants interacted directly with the prototype and had to simulate a prostate MRI diagnosis using the prototype and provided MRI images, following a specific scenario. They were encouraged to vocalize their thoughts throughout the process using the think-aloud method. ⁹⁷ To capture these thoughts, all sessions were recorded, transcribed, and analyzed using a thematic analysis approach ⁹⁸ to identify key topics, issues, and design implications raised by participants. The participants’ feedback was helpful to refine the prototypes iteratively. Some design ideas that emerged through the evaluation were already addressed within iterative design cycles; other potential features not implemented in the prototype were discussed. Participants shared insights on the benefits and drawbacks of the features. Familiarity with the prototype likely facilitated participants’ ability to envision the planned features as potential additions to the software solution.

5.1 Low-fidelity prototypes

The Lo-Fi prototypes were generated from our workshops and through internal sketching efforts within the research team. ⁹⁹ The interdisciplinary collaboration in our study fostered a rich exchange of ideas, ranging from user-centric interface designs to interactive features for the prototypes aimed at tackling the identified problems, while our focus remained on exploring how a potential solution could best meet user needs and on uncovering crucial factors that must be considered ¹⁰⁰ when designing for actual work practices.

5.1.1 Online design thinking workshop

During the workshop, we expanded upon the user needs and requirements that we gathered and created POV statements to aid in specifying the particular requirements, and employed HMW statements to discover several concepts centered around the intended results. For example, while working on the requirement “segmentation of distinct prostate regions” (reference to requirement #2), we crafted the POV statement “The radiologist needs the segmentation of the individual zones of the prostate for the biopsies because it is time-consuming to annotate them by hand.” which led to the HMW statement of “How might we provide the segmentation of the individual zones of the prostate (central gland, peripheral zone, transitional zone, seminal vesicle) for the biopsies to save time?” which was addressed through the prototype that was designed during the workshop.

Through collaborative efforts within two interdisciplinary groups, we successfully crafted two Lo-Fi prototypes. Subsequently, we fine-tuned certain elements of these prototypes, guided by the feedback received during the workshop. To enhance the comprehensiveness of our designs, we incorporated sticky notes containing some HMW statements, insights, and ideas, addressing aspects that couldn’t be accommodated within the prototypes due to time constraints. This process ensures that the final prototypes not only reflect the diverse perspectives within the groups but also integrate valuable insights for a more refined outcome.

Figure 2 presents the prototype created by group one during the workshop. Their primary focus was on consolidating crucial information into a single frame, minimizing the need for extensive interaction with various screens and elements to reduce distractions and time. According to the requirements list (see Subsection 2.2 to check the requirements list), the participants demonstrated the visualization of AI-detected zones, lesions, and segmentation through contours (reference to requirements #1, #2, #3). Emphasizing the importance of displaying key information about lesions and the prostate, they showcased overlays. A table was designed to provide an overview of all AI-identified lesions, including the PI-RADS score (requirement #4), with options for result verification and explanations of the score. Also, a detailed view featured extensive information about lesions and patients, offering an option for editing. Participants also streamlined the reporting process, incorporating relevant images directly into the report through checkboxes activated by the “Generate Report” button (requirement #6). However, time constraints hindered the completion of a graphical representation of lesion progress.

Figure 2:

Low-fidelity prototype by Group 1 from online design thinking workshop.

In the second prototype presented in Figure 3 made by group two, the radiologist in the group identified comparing AI-based pre-diagnoses with original MRI images as an important user need, which was then focused on. The radiologists’ design idea emphasizes a parallel order of the AI results (“annotated images”) and the original MRI images (“original images”). These images should be linked to each other and change synchronously while scrolling through the mouse without additional interaction elements. Also, the detection of the prostate borders as well as the segmentation of different zones and lesions were visualized through contours or a heatmap (requirements #1, #2, #3). Participants incorporated an option for corrections of the markings. The right side of the interface shows space for reporting (requirement #6). In a table format, the AI should provide automatic calculations of the prostate and lesions volume (requirement #5). It is intended to ensure the classification of identified lesions following the PI-RADS scheme and to automatically calculate the PI-RADS score (requirement #4). Other HMW statements, which couldn’t be addressed in the prototype, refer to a Prostate Sector Map according to PI-RADS, which should be provided and generated automatically, and show the individual lesions, their weighting zone, and the PI-RADS score.

Figure 3:

Low-fidelity prototype by Group 2 from online design thinking workshop.

5.1.2 On-site design thinking workshop

A second DT workshop was conducted on-site with 20 participants divided into four groups. Consequently, four paper prototypes were crafted as seen in Figure 4, reflecting the collaborative efforts of the groups and demonstrating the convergence of diverse perspectives into cohesive design solutions for a specific problem.

Figure 4:

Low-fidelity paper prototypes from Group 1 to Group 4 from on-site design thinking workshop. (a) Paper prototype of Group 1. (b) Paper prototype of Group 2. (c) Paper prototype of Group 3. (d) Paper prototype of Group 4.

Group 1 concentrated on showcasing the automatic identification and segmentation of distinct prostate regions using contours, offering a 3D view, and providing an option for radiologist verification (requirements #1 and #2). Group 2 focused on visualizing the automatic detection and localization of potential lesions with classification, emphasizing its potential role in identifying cancer severity among patients and prioritizing cases beforehand (requirements #3 and #4). Group 3 envisioned the automatic calculation of prostate volume and PSA density, emphasizing the importance of allowing radiologists to adjust AI-detected borders for accuracy (requirement #5). Group 4 demonstrated a standardized reporting system that not only provided feedback on AI results but also included a Likert scale option for radiologists to indicate the severity of a misdiagnosis by AI (requirements #6 and #7).

5.1.3 Low-fidelity prototype using figma

Following the prior prototypes and feedback, our research team used quick visual sketching in Figma to support the creative ideation process, ⁹⁹ which allowed for rapid idea externalization, including the use of images like MRI scans (Figure 5), and proved pivotal in creating a well-informed, user-centered design.

Figure 5:

Low-fidelity prototype: figma sketch of diagnostic tool UI with the PI-RADS sector map ¹⁰¹ and a T2W image.

Since our main target users are German radiologists, the UI text for this Lo-Fi and subsequent prototypes was in German to ensure a realistic and familiar user experience. All of the seven Lo-Fi prototypes were discussed and refined within the research team and used for the next iteration of prototyping to create a Hi-Fi prototype.

5.2 High-fidelity prototype

The primary goal in translating the Lo-Fi prototypes into a Hi-Fi prototype was to create a user experience that closely resembled interacting with a finished product. To enhance the realism of the prototype interaction, real MRI examination data and actual AI-generated output were utilized and integrated into the Hi-Fi prototype, which was implemented as a front-end web application using AngularJS. The AI developers of our team were concurrently implementing and training the AI algorithm. They used the Semi-Supervised Learning (SSL) technique to train our AI model, since it produces more accurate outcomes by learning from both labeled and unlabeled data, ^{102 , 103} and has the potential for medical image segmentation. ¹⁰⁴ The technical details of the implementation and training process of the AI algorithm are beyond the scope of this paper. Although the algorithm was not directly linked to the back-end of the Hi-Fi prototype, the AI results were visibly presented within the UI using Adobe Photoshop.^[8]

Once the AI algorithms have processed the MRI images, the results are made available via a web-based front-end application, allowing users to examine each image in greater detail. The AI primarily focuses on detecting clinically significant lesions and accurately delineating both the whole prostate and the lesions, which are presented to the user in a simplified DICOM viewer. Additionally, the solution generates a structured and informative graphic to be included in the radiological report.

Here, we will provide a detailed description of our UI and highlight key features. At first, radiologists start to select an MRI exam from an overview page listing all exams (scans) processed with patient details, exam date (newest first), and the AI’s processing status. The details page, as shown in Figure 6, is split into a left side for textual data and a right-side viewer for graphical data, such as the MRI images and lesion graphics. This simultaneous display allows users to easily make connections and comparisons, which in turn reduces cognitive load by eliminating manual view switching.

Figure 6:

High-fidelity prototype: detailed UI with examination-details page of a patient.

The viewer includes the MRI series relevant to PI-RADS scoring. Users can navigate between series via top buttons, the scroll bar, or direct mouse-wheel scrolling, a common interaction element in the DICOM viewer and explicitly identified as a user need during the workshop. The viewer also displays AI-generated delineations. Through mask options, the prostate segmentation, PZ, and any lesion detected are presented and classified as significant by the AI algorithm (Figure 7). Activated masks remain visible throughout the image stack during scrolling. For orientation, the scroll bar thumbnails were color-coded according to active masks, with distinct colors differentiating organ/zone from lesion delineations.

Figure 7:

High-fidelity prototype: viewer options with four different kinds of segmentation masks. (a) Prostate. (b) Peripheral zone (PZ). (c) Lesion. (d) Prostate, PZ, and Lesion.

Data transfer features are included (copy to clipboard, download, or export to Picture Archiving and Communication System (PACS)), enabling users to view edited series in their DICOM viewer or integrate them directly into reports. The viewer can display the AI-generated structured graphic (Figure 8), reducing input errors, increasing efficiency, and ensuring consistent outputs transferable to final reports.

Figure 8:

High-fidelity prototype: AI-generated sector map with structured table.

In Figure 9, the inputs are structured into sections aligned with the diagnostic workflow observed in our context study, reducing cognitive effort by maintaining familiar procedures. Checkboxes in each section support process documentation and help radiologists resume work after interruptions. The first section addresses diagnostic image quality, which is pre-filled by AI according to the concept. The prostate measurements section shows PSA, prostate volume, and PSA density, automatically calculated from AI segmentation to improve efficiency and reduce errors. The option to validate the values is available by assessing the right-side images. The lesion-finding section lists AI-detected lesions, which remain editable by the user. While the PI-RADS score is automatically derived from registered values, users can input a manual score for ambiguous cases. The lesion-finding section allows checking corresponding delineations, adding new lesions, or hiding registered ones to manage graphics and dismiss AI suggestions. The documentation section covers the local and external cancer spreading and additional findings, using structured fields (oriented on the PI-RADS v2.1 benign findings list ⁴³ ). The final part registers the overall assessment score. Each section is structured as a table where the current exam is a row.

Figure 9:

High-fidelity prototype: detailed UI with complete input form.

5.3 Middle-fidelity prototype

Within our iterative process, multiple design and evaluation cycles were conducted. After evaluating the Hi-Fi version, a Mid-Fi prototype was created in Figma (Figure 10).

Figure 10:

Middle-fidelity prototype after iteration: diagnostic tool UI with a more streamlined design.

The Mid-Fi prototype aims to reduce perceptual and cognitive load by enabling radiologists to make quick, pre-defined decisions on AI-suggested abnormalities. This reflects the radiologists’ request to mark easily categorizable lesions using their expertise (P09). The AI suggestions can validate decisions or be overridden in case of disagreement, serving purely as supportive augmentation. Confidence bars promote diagnostic caution, and an on/off switch enables AI use either as pre-diagnostic guidance or as a confirmatory tool.

6.1 Usability evaluation: first iteration

The Hi-Fi prototype was evaluated first, with radiologists generally endorsing its overall concept during the study. P06 highlighted his recognition by saying “I would take it how it is right now, with the points we discussed, the changes, improvements, etc. I would take it, let it be certified, and bring it to the market.” The UI was perceived as easy to use, enabling users to orient themselves quickly; however, this applies mainly to individuals “who are familiar with the subject matter” (P02).

The presentation of prostate delineations and identified lesions on the T2W images was considered beneficial by all participants, enabling users to examine the AI algorithm’s outcomes easily. Highlighting suspicious regions facilitated the assessment of lesion number, size, and location, while navigation through the full image stack supported a comprehensive understanding of AI conclusions. Active mask visualization “[.] facilitates and fastens a quick overview. If I want to see where the algorithm finds something, I don’t need to scroll through the whole image set but can approach it directly” (P05).

Participants acknowledged the importance of the prostate measurement section and the relevance of each displayed value for the graphic. Especially, the automated calculation of prostate volume and PSA density was valued. As P06 noted, “This is in principle what facilitates one’s work, that you don’t have to type into your calculator or ask Siri.”

Incorporating the PI-RADS score for each sequence, as well as the overall final score, was considered essential for the assisted workflow. The automated calculation of the final lesion score and the option for manual input were regarded as practical. The potential inclusion of visual alerts in cases where the manually assigned score contradicts the PI-RADS algorithm was regarded as a beneficial safeguard.

During task completion, participants used separate screens for the prototype and their DICOM viewer, as intended. The concept of assessing images within the DICOM viewer while simultaneously utilizing the AI solution on a distinct screen was deemed logical. According to P07, he does the diagnosis in the DICOM viewer anyway, and then he cross-checks on the other side. Conversely, P05 highlighted the potential advantages of direct integration into the existing PACS, which could reduce cognitive load.

The input form on the left side received a unanimous positive response from all participants, appreciating the structure’s alignment with their current workflow. This arrangement allowed the participants to methodically progress through individual steps and seamlessly proceed once they were completed. P07 said, “This already represents my approach very nicely, how I would work through the diagnosis. That’s what I need. That I can go through this and mentally tick my things. And then I have my finished graphic and report.” All radiologists (also in the second evaluation cycle) appreciated automatic reporting, valuing any functionality that saves time and reduces workload. The ability to manually add lesions and correct values is deemed beneficial. However, P02 mentions a desire to have more flexibility when selecting secondary findings, e.g., by hiding irrelevant options. As this is a minor issue, we were able to change that directly in the Hi-Fi prototype.

The prostate sector map and the marked lesions also received positive feedback, along with suggestions for improvement. Especially, marking lesions based on their delineation results in a more precise representation was acknowledged. All participants recognized the advantages of automatically generating a lesion graphic, such as its time-saving and enhanced efficiency. Furthermore, it eradicates the potential for transfer errors and guarantees a consistent graphic outcome.

Although participants agreed on most features, variations exist between the values, preferences, and workflows of individual radiologists. Divergences emerged regarding aspects such as the importance of the lesion volume, information included in the report graphic, and the use of color for highlighting. This represents the challenge that design decisions valued by some can be inconvenient for others. As suggested by participants, customizable settings could mitigate this issue.

6.2 Usability evaluation: second iteration

In our case, the final version is the Mid-Fi prototype, which was created after evaluating the Hi-Fi prototype. In general, the AI is viewed as effective not only when making accurate assessments but also for not significantly disrupting the workflow when predictions are incorrect. While P08 expects high accuracy, he recognizes false positives as inevitable and suggests a threshold to limit displayed lesions, though he wants to see enough abnormalities to avoid missing cancerous lesions. P04 and P06 readily override AI assessments, appreciating that the prototype supports human oversight to compensate for AI fallibility. P07 highlighted the importance of quickly and intuitively dismissing false positives, valuing the ability to easily delete AI suggestions, particularly in the TZ where they are frequent. P07 found it more efficient to correct an inaccurate AI suggestion with minimal adjustments than to manually input findings from scratch. Though he acknowledged the risk of diminishing intuition, which could be particularly problematic given the relatively frequent occurrence of edge cases.

P09 appreciates the AI by saying, “Of course, I think it is good that the system shows you where an abnormal finding is and allows you to see boundaries, so I can get an idea of how large the finding is”. He also appreciated the AI-suggested lesions in the list, praising the clear presentation of their details, such as numbering, zone assignment, and size.

The confidence bars were positively received as they offer a more natural experience (P08) and allow participants to choose among multiple AI suggestions rather than relying on a single prediction. P04 compared this to differential diagnosis, where physicians weigh several possibilities before deciding, and P06 highlighted its novelty for prostate cancer, “I find it helpful. [It] is already used in other fields [.], where diagnostic suggestions are provided along with a probability. I think it is an interesting feature. I have never seen anything like this in prostate cancer diagnosis before”.

As P09 evaluated a pre-version of the Mid-Fi prototype, he found it challenging to link lesions in the MRI sequence to the corresponding AI-generated entries in the list, particularly when multiple lesions were present, “On the left, I have two lesions labeled as ‘Lesion 1’ and ‘Lesion 2.’ When I see the red-highlighted lesion in the MRI image on the right, I cannot tell which of the two it is” (P09). He further suggested sorting AI-identified lesions by size, from largest to smallest, reflecting his typical workflow in which larger lesions are examined first before identifying smaller ones. That was already implemented in the final version, which can be seen in Figure 10.

6.3 Design recommendations for redesign

During the evaluation, several design recommendations for improvement were suggested. While we will address key points here, we will not delve into suggestions for minor changes, as they are not the primary focus.

Similar to other AI-generated data, the ability to modify map markings should always be adjustable to integrate smoothly into clinical workflows. This includes the ability to remove or add lesions and, as P06 and P07 proposed, to manually delineate new ones directly on the MRI. This would allow the system to auto-populate values based on the image data within the delineated region. We believe that incorporating this feature may empower radiologists to have more control over the mapping process, while also addressing concerns regarding the practicality and efficiency of manual lesion delineation expressed by P05.

The evaluation revealed the importance of preserving overwritten AI-generated results while ensuring clear visual differentiation. Retaining original outputs allows radiologists to compare their input with AI suggestions, track discrepancies over time, and monitor AI performance. To support this, a dedicated dashboard might facilitate this monitoring. Visual contrast between rectified and AI-generated values, such as overlaying both delineations, as suggested by P07, was considered essential to avoid confusion and make discrepancies transparent. This differentiation aids radiologists in indicating AI-generated results and understanding the process leading to outcomes. Moreover, it facilitates crafting reports, as discrepancies from AI output can be communicated to the referrer. Moreover, maintaining access to original values supports retroactive traceability, particularly for cases requiring re-evaluation.

For the Hi-Fi prototype, concerns were raised regarding the accuracy of markings on the prostate sector map. “The graphic does not match the anatomy” (P02). In this case, the lesion would have to be corrected manually. Participants further expected large lesions to be consistently marked across all relevant planes (base, middle, apex) and criticized the absence of markings in sagittal and coronal views or alternative sequences (e.g., DWI), though opinions on their relevance varied. P02 suggested color-coding lesions (“[.] to distinguish between them, with the target lesion in red and secondary lesions in a different color”). While this was then implemented in the Mid-Fi prototype, P07 and P08 misinterpreted the colors as indicators of cancer severity, which might be an unintended but potentially useful feature, as such coding is familiar from other software.

7.1 Limitations

While our contribution offers valuable insights into our research goal, it is important to acknowledge its limitations, particularly concerning our sample. Although we have invested great effort in recruiting radiologists, we faced considerable challenges in the process due to the scarcity of available professionals. Nevertheless, we were able to engage experienced individuals in our workshops and evaluation sessions, providing valuable insights.

Appropriation phase in a Design Case Study (DCS) primarily focuses on the long-term adoption of the technical artifact. However, due to our current stage, we have not yet implemented the new tool within radiological organizations over an extended period. Moving forward, we plan to integrate the final AI solution into real-world settings and incorporate radiologists’ feedback through a human-in-the-loop approach, necessitating further evaluation and appropriateness studies with larger participant pools and long-term evaluations.

An additional constraint concerned our DT workshops, which had to be conducted within a few hours due to participants’ limited availability. Therefore, we used an adapted design sprint with DT as the methodological framework and, instead of requiring participants to conduct their own research, provided them with insights from our prior empirical study. Radiologists explained their daily practices and challenges to non-medical participants in the first workshop, and we shared these insights with students in the second workshop to help them empathize with real users. However, some HCI methods proved difficult for radiologists to grasp, leading to critical discussions about the purpose of DT and its structured approach. With little time to bridge these knowledge gaps, it was also challenging for us as researchers to convey the value of HCD, highlighting the need for more time on foundational understanding in future workshops.