Trust and stakeholder perspectives on the implementation of AI tools in clinical radiology

Mapping and Summarizing the Research on AI Systems for Automating Medical History Taking and Triage: Scoping Review

BACKGROUND

The integration of artificial intelligence (AI) systems for automating medical history taking and triage can significantly enhance patient flow in health care systems. Despite the promising performance of numerous AI studies, only a limited number of these systems have been successfully integrated into routine health care practice. To elucidate how AI systems can create value in this context, it is crucial to identify the current state of knowledge, including the readiness of these systems, the facilitators of and barriers to their implementation, and the perspectives of various stakeholders involved in their development and deployment.

OBJECTIVE

This study aims to map and summarize empirical research on AI systems designed for automating medical history taking and triage in health care settings.

METHODS

The study was conducted following the framework proposed by Arksey and O'Malley and adhered to the PRISMA-ScR (Preferred Reporting Items for Systematic reviews and Meta-Analyses extension for Scoping Reviews) guidelines. A comprehensive search of 5 databases-PubMed, CINAHL, PsycINFO, Scopus, and Web of Science-was performed. A detailed protocol was established before the review to ensure methodological rigor.

RESULTS

A total of 1248 research publications were identified and screened. Of these, 86 (6.89%) met the eligibility criteria. Notably, most (n=63, 73%) studies were published between 2020 and 2022, with a significant concentration on emergency care (n=32, 37%). Other clinical contexts included radiology (n=12, 14%) and primary care (n=6, 7%). Many (n=15, 17%) studies did not specify a clinical context. Most (n=31, 36%) studies used retrospective designs, while others (n=34, 40%) did not specify their methodologies. The predominant type of AI system identified was the hybrid model (n=68, 79%), with forecasting (n=40, 47%) and recognition (n=36, 42%) being the most common tasks performed. While most (n=70, 81%) studies included patient populations, only 1 (1%) study investigated patients' views on AI-based medical history taking and triage, and 2 (2%) studies considered health care professionals' perspectives. Furthermore, only 6 (7%) studies validated or demonstrated AI systems in relevant clinical settings through real-time model testing, workflow implementation, clinical outcome evaluation, or integration into practice. Most (n=76, 88%) studies were concerned with the prototyping, development, or validation of AI systems. In total, 4 (5%) studies were reviews of several empirical studies conducted in different clinical settings. The facilitators and barriers to AI system implementation were categorized into 4 themes: technical aspects, contextual and cultural considerations, end-user engagement, and evaluation processes.

CONCLUSIONS

This review highlights current trends, stakeholder perspectives, stages of innovation development, and key influencing factors related to implementing AI systems in health care. The identified literature gaps regarding stakeholder perspectives and the limited research on AI systems for automating medical history taking and triage indicate significant opportunities for further investigation and development in this evolving field.

Clinicians' perspectives on the use of artificial intelligence to triage MRI brain scans

Artificial intelligence (AI) tools can triage radiology scans to streamline the patient pathway and also relieve clinician workload. Validated AI tools can mitigate the delays in reporting scans by flagging time-sensitive and actionable findings. In this study, we aim to investigate current stakeholder perspectives and identify obstacles to integrating AI in clinical pathways. We created a survey to ascertain the perspectives of 133 clinicians across the United Kingdom regarding the acceptability of an AI tool that triages MRI brain scans into 'normal' and 'abnormal'. As part of this survey, we supplied clinicians with information on training and validation case numbers, model performance, validation using unseen data, and explainability saliency maps. With regards to the specific use case of AI in MRI brain scans, 71% of respondents preferred the use of an AI-assisted triage compared to the current system without triage, typically chronologically. Notably, information that explained and helped visualise the AI model's decision making was found to improve clinician confidence. When shown a heatmap, 60% of participants felt more confident in the AI's decision. The results of this short communication demonstrate a positive support for the implementation of AI-assistive tools in triage.

Enhancing Diagnostic Accuracy of Lung Nodules in Chest Computed Tomography Using Artificial Intelligence: Retrospective Analysis

BACKGROUND

Uncertainty in the diagnosis of lung nodules is a challenge for both patients and physicians. Artificial intelligence (AI) systems are increasingly being integrated into medical imaging to assist diagnostic procedures. However, the accuracy of AI systems in identifying and measuring lung nodules on chest computed tomography (CT) scans remains unclear, which requires further evaluation.

OBJECTIVE

This study aimed to evaluate the impact of an AI-assisted diagnostic system on the diagnostic efficiency of radiologists. It specifically examined the report modification rates and missed and misdiagnosed rates of junior radiologists with and without AI assistance.

METHODS

We obtained effective data from 12,889 patients in 2 tertiary hospitals in Beijing before and after the implementation of the AI system, covering the period from April 2018 to March 2022. Diagnostic reports written by both junior and senior radiologists were included in each case. Using reports by senior radiologists as a reference, we compared the modification rates of reports written by junior radiologists with and without AI assistance. We further evaluated alterations in lung nodule detection capability over 3 years after the integration of the AI system. Evaluation metrics of this study include lung nodule detection rate, accuracy, false negative rate, false positive rate, and positive predictive value. The statistical analyses included descriptive statistics and chi-square, Cochran-Armitage, and Mann-Kendall tests.

RESULTS

The AI system was implemented in Beijing Anzhen Hospital (Hospital A) in January 2019 and Tsinghua Changgung Hospital (Hospital C) in June 2021. The modification rate of diagnostic reports in the detection of lung nodules increased from 4.73% to 7.23% (χ21=12.15; P<.001) at Hospital A. In terms of lung nodule detection rates postimplementation, Hospital C increased from 46.19% to 53.45% (χ21=25.48; P<.001) and Hospital A increased from 39.29% to 55.22% (χ21=122.55; P<.001). At Hospital A, the false negative rate decreased from 8.4% to 5.16% (χ21=9.85; P=.002), while the false positive rate increased from 2.36% to 9.77% (χ21=53.48; P<.001). The detection accuracy demonstrated a decrease from 93.33% to 92.23% for Hospital A and from 95.27% to 92.77% for Hospital C. Regarding the changes in lung nodule detection capability over a 3-year period following the integration of the AI system, the detection rates for lung nodules exhibited a modest increase from 54.6% to 55.84%, while the overall accuracy demonstrated a slight improvement from 92.79% to 93.92%.

CONCLUSIONS

The AI system enhanced lung nodule detection, offering the possibility of earlier disease identification and timely intervention. Nevertheless, the initial reduction in accuracy underscores the need for standardized diagnostic criteria and comprehensive training for radiologists to maximize the effectiveness of AI-enabled diagnostic systems.

Bridging the Gap: A Mixed-Methods Study on Factors Influencing Breast Cancer Clinicians' Decisions to Use Clinical Prediction Models

Background. Clinical prediction models provide tailored risk estimates that can help guide decisions in breast cancer care. Despite their potential, few models are widely used in clinical practice. We aimed to identify the factors influencing breast cancer clinicians' decisions to adopt prediction models and assess their relative importance. Methods. We conducted a mixed-methods study, beginning with semi-structured interviews, followed by a nationwide online survey. Thematic analysis was used to qualitatively summarize the interviews and identify key factors. For the survey, we used descriptive analysis to characterize the sample and Mann-Whitney U and Kruskal-Wallis tests to explore differences in score (0 = not important to 10 = very important) distributions. Results. Interviews (N = 16) identified eight key factors influencing model use. Practical/methodological factors included accessibility, cost, understandability, objective accuracy, actionability, and clinical relevance. Perceptual factors included acceptability, subjective accuracy, and risk communication. In the survey (N = 146; 137 model users), clinicians ranked online accessibility (median score = 9 [interquartile range = 8-10]) as most important. Cost was also highly rated, with preferences for freely available models (9 [8-10]) and those with reimbursable tests (8 [8-10]). Formal regulatory approval (7 [5-8]) and direct integration with electronic health records (6 [3-8]) were considered less critical. Subgroup analysis revealed differences in score distributions; for example, clinicians from general hospitals prioritized inclusion of new biomarkers more than those in academic settings. Conclusions. Breast cancer clinicians' decisions to initiate use of prediction models are influenced by practical and perceptual factors, extending beyond technical metrics such as discrimination and calibration. Addressing these factors more holistically through collaborative efforts between model developers, clinicians, and communication and implementation experts, for instance, by developing clinician-friendly online tools that prioritize usability and local adaptability, could increase model uptake.

Highlights

Accessibility, cost, and practical considerations, such as ease of use and clinical utility, were prioritized slightly more than technical validation metrics, such as discrimination and calibration, when deciding to start using a clinical prediction model.Most breast cancer clinicians valued models with clear inputs (e.g., variable definitions, cutoffs) and outputs; few were interested in the exact model specifications.Perceptual or subjective factors, including perceived accuracy and peer acceptability, also influenced model adoption but were secondary to practical considerations.Sociodemographic variables, such as clinical specialization and hospital setting, influenced the importance of factors for model use.

Radiation Oncology Resident Education: Is Change Needed?

Leading successful change efforts first requires assessment of the "before change" environment and culture. At our institution, the radiation oncology (RO) residents follow a longitudinal didactic learning program consisting of weekly 1-h lectures, case conferences, and journal clubs. The resident didactic education series format has not changed since its inception over 10 years ago. We evaluated the perceptions of current residents and faculty about the effectiveness of the curriculum in its present form. Two parallel surveys were designed, one each for residents and attendings, to assess current attitudes regarding the effectiveness and need for change in the RO residency curriculum, specifically the traditional didactic lectures, the journal club sessions, and the case conferences. We also investigated perceived levels of engagement among residents and faculty, whether self-assessments would be useful to increase material retention, and how often the content of didactic lectures is updated. Surveys were distributed individually to each resident (N = 10) and attending (N = 24) either in-person or via Zoom. Following completion of the survey, respondents were informally interviewed about their perspectives on the curriculum's strengths and weaknesses. Compared to 46% of attendings, 80% of RO residents believed that the curriculum should be changed. Twenty percent of residents felt that the traditional didactic lectures were effective in preparing them to manage patients in the clinic, compared to 74% of attendings. Similarly, 10% of residents felt that the journal club sessions were effective vs. 42% of attendings. Finally, 40% of residents felt that the case conferences were effective vs. 67% of attendings. Overall, most respondents (56%) favored change in the curriculum. Our results suggest that the perceptions of the residents did not align with those of the attending physicians with respect to the effectiveness of the curriculum and the need for change. The discrepancies between resident and faculty views highlight the importance of a dedicated change management effort to mitigate this gap. Based on this project, we plan to propose recommended changes in structure to the residency program directors. Main changes would be to increase the interactive nature of the course material, incorporate more ways to increase faculty engagement, and consider self-assessment questions to promote retention. Once we get approval from the residency program leadership, we will follow Kotter's "Eight steps to transforming your organization" to ensure the highest potential for faculty to accept the expectations of a new curriculum.

The American Society of Radiologic Technologists (ASRT) AI educator survey: A cross-sectional study to explore knowledge, experience, and use of AI within education

INTRODUCTION

Artificial Intelligence (AI) is revolutionizing medical imaging and radiation therapy. AI-powered applications are being deployed to aid Medical Radiation Technologists (MRTs) in clinical workflows, decision-making, dose optimisation, and a wide range of other tasks. Exploring the levels of AI education provided across the United States is crucial to prepare future graduates to deliver the digital future. This study aims to assess educators' levels of AI knowledge, the current state of AI educational provisions, the perceived challenges around AI education, and important factors for future advancements.

METHODS

An online survey was electronically administered to all radiologic technologists in the American Society of Radiologic Technologists (ASRT) database who indicated that they had an educator role in the United States. This was distributed through the membership of the ASRT, from February to April 2023. All quantitative data was analysed using frequency and descriptive statistics. The survey's open-ended questions were analysed using a conceptual content analysis approach.

RESULTS

Out of 5,066 educators in the ASRT database, 373 valid responses were received, resulting in a response rate of 7.4%. Despite 84.5% of educators expressing the importance of teaching AI, 23.7% currently included AI in academic curricula. Of the 76.3% that did not include AI in their curricula, lack of AI knowledge among educators was the top reason for not integrating AI in education (59.1%). Similarly, AI-enabled tools were utilised by only 11.1% of the programs to assist teaching. The levels of trust in AI varied among educators.

CONCLUSION

The study found that although US educators of MRTs have a good baseline knowledge of general concepts regarding AI, they could improve on the teaching and use of AI in their curricula. AI training and guidance, adequate time to develop educational resources, and funding and support from higher education institutions were key priorities as highlighted by educators.

Deep learning-assisted distinguishing breast phyllodes tumours from fibroadenomas based on ultrasound images: a diagnostic study

OBJECTIVES

To evaluate the performance of ultrasound-based deep learning (DL) models in distinguishing breast phyllodes tumours (PTs) from fibroadenomas (FAs) and their clinical utility in assisting radiologists with varying diagnostic experiences.

METHODS

We retrospectively collected 1180 ultrasound images from 539 patients (247 PTs and 292 FAs). Five DL network models with different structures were trained and validated using nodule regions annotated by radiologists on breast ultrasound images. DL models were trained using the methods of transfer learning and 3-fold cross-validation. The model demonstrated the best evaluation index in the 3-fold cross-validation was selected for comparison with radiologists' diagnostic decisions. Two-round reader studies were conducted to investigate the value of DL model in assisting 6 radiologists with different levels of experience.

RESULTS

Upon testing, Xception model demonstrated the best diagnostic performance (area under the receiver-operating characteristic curve: 0.87; 95% CI, 0.81-0.92), outperforming all radiologists (all P < .05). Additionally, the DL model enhanced the diagnostic performance of radiologists. Accuracy demonstrated improvements of 4%, 4%, and 3% for senior, intermediate, and junior radiologists, respectively.

CONCLUSIONS

The DL models showed superior predictive abilities compared to experienced radiologists in distinguishing breast PTs from FAs. Utilizing the model led to improved efficiency and diagnostic performance for radiologists with different levels of experience (6-25 years of work).

ADVANCES IN KNOWLEDGE

We developed and validated a DL model based on the largest available dataset to assist in diagnosing PTs. This model has the potential to allow radiologists to discriminate 2 types of breast tumours which are challenging to identify with precision and accuracy, and subsequently to make more informed decisions about surgical plans.

Prediction of neurologic outcome after out-of-hospital cardiac arrest: An interpretable approach with machine learning

Out-of-hospital cardiac arrest (OHCA) is a critical condition with low survival rates. In patients with a return of spontaneous circulation, brain injury is a leading cause of death. In this study, we propose an interpretable machine learning approach for predicting neurologic outcome after OHCA, using information available at the time of hospital admission.

METHODS

The study population were 55 615 OHCA cases registered in the Swedish Cardiopulmonary Resuscitation Registry between 2010 and 2020. The dataset was split to training and validation sets (for model development) and test set (for evaluation of the final model). We used an XGBoost algorithm with stratified, repeated 10-fold cross-validation along with Optuna framework for hyperparameters tuning. The final model was trained on 10 features selected based on the importance scores and evaluated on the test set in terms of discrimination, calibration and bias-variance tradeoff. We used SHapley Additive exPlanations to address the 'black-box' model and align with eXplainable artificial intelligence.

RESULTS

The final model achieved: area under the receiver operating characteristic value 0.964 (95% confidence interval (CI) [0.960-0.968]), sensitivity 0.606 (95% CI [0.573-0.634]), specificity 0.975 (95% CI [0.972-0.978]), positive predictive value (PPV) 0.664 (95% CI [0.625-0.696]), negative predictive value (NPV) 0.969 (95% CI [0.966-0.972]), macro F1 0.803 (95% CI [0.788-0.816]), and showed a very good calibration. SHAP features with the highest impact on the model's output were:'ROSC on arrival to hospital', 'Initial rhythm asystole' and 'Conscious on arrival to hospital'.

CONCLUSIONS

The XGBoost machine learning model with 10 features available at the time of hospital admission showed good performance for predicting neurologic outcome after OHCA, with no apparent signs of overfitting.

Expanding role and scope of artificial intelligence in the field of gastrointestinal pathology

Digital pathology (DP) and its subsidiaries including artificial intelligence (AI) are rapidly making inroads into the area of diagnostic anatomic pathology (AP) including gastrointestinal (GI) pathology. It is poised to revolutionize the field of diagnostic AP. Historically, AP has been slow to adopt digital technology, but this is changing rapidly, with many centers worldwide transitioning to DP. Coupled with advanced techniques of AI such as deep learning and machine learning, DP is likely to transform histopathology from a subjective field to an objective, efficient, and transparent discipline. AI is increasingly integrated into GI pathology, offering numerous advancements and improvements in overall diagnostic accuracy, efficiency, and patient care. Specifically, AI in GI pathology enhances diagnostic accuracy, streamlines workflows, provides predictive insights, integrates multimodal data, supports research, and aids in education and training, ultimately improving patient care and outcomes. This review summarized the latest developments in the role and scope of AI in AP with a focus on GI pathology. The main aim was to provide updates and create awareness among the pathology community.

Bridging the Gap: Challenges and Strategies for the Implementation of Artificial Intelligence-based Clinical Decision Support Systems in Clinical Practice

OBJECTIVES

Despite the surge in development of artificial intelligence (AI) algorithms to support clinical decision-making, few of these algorithms are used in practice. We reviewed recent literature on clinical deployment of AI-based clinical decision support systems (AI-CDSS), and assessed the maturity of AI-CDSS implementation research. We also aimed to compare and contrast implementation of rule-based CDSS with implementation of AI-CDSS, and to give recommendations for future research in this area.

METHODS

We searched PubMed and Scopus for publications in 2022 and 2023 that focused on AI and/or CDSS, health care, and implementation research, and extracted: clinical setting; clinical task; translational research phase; study design; participants; implementation theory, model or framework used; and key findings.

RESULTS

We selected and described a total of 31 recent papers addressing implementation of AI-CDSS in clinical practice, categorised into four groups: (i) Implementation theories, frameworks, and models (4 papers); (ii) Stakeholder perspectives (22 papers); (iii) Implementation feasibility (three papers); and (iv) Technical infrastructure (2 papers). Stakeholders saw potential benefits of AI-CDSS, but emphasized the need for a strong evidence base and indicated that systems should fit into clinical workflows. There were clear similarities with rule-based CDSS, but also differences with respect to trust and transparency, knowledge, intellectual property, and regulation.

CONCLUSIONS

The field of AI-CDSS implementation research is still in its infancy. It can be strengthened by grounding studies in established theories, models and frameworks from implementation science, focusing on the perspectives of stakeholder groups other than healthcare professionals, conducting more real-world implementation feasibility studies, and through development of reusable technical infrastructure that facilitates rapid deployment of AI-CDSS in clinical practice.

Clinical, Cultural, Computational, and Regulatory Considerations to Deploy AI in Radiology: Perspectives of RSNA and MICCAI Experts

The Radiological Society of North of America (RSNA) and the Medical Image Computing and Computer Assisted Intervention (MICCAI) Society have led a series of joint panels and seminars focused on the present impact and future directions of artificial intelligence (AI) in radiology. These conversations have collected viewpoints from multidisciplinary experts in radiology, medical imaging, and machine learning on the current clinical penetration of AI technology in radiology and how it is impacted by trust, reproducibility, explainability, and accountability. The collective points-both practical and philosophical-define the cultural changes for radiologists and AI scientists working together and describe the challenges ahead for AI technologies to meet broad approval. This article presents the perspectives of experts from MICCAI and RSNA on the clinical, cultural, computational, and regulatory considerations-coupled with recommended reading materials-essential to adopt AI technology successfully in radiology and, more generally, in clinical practice. The report emphasizes the importance of collaboration to improve clinical deployment, highlights the need to integrate clinical and medical imaging data, and introduces strategies to ensure smooth and incentivized integration. Keywords: Adults and Pediatrics, Computer Applications-General (Informatics), Diagnosis, Prognosis © RSNA, 2024.

Implications of Large Language Models for Quality and Efficiency of Neurologic Care: Emerging Issues in Neurology

Large language models (LLMs) are advanced artificial intelligence (AI) systems that excel in recognizing and generating human-like language, possibly serving as valuable tools for neurology-related information tasks. Although LLMs have shown remarkable potential in various areas, their performance in the dynamic environment of daily clinical practice remains uncertain. This article outlines multiple limitations and challenges of using LLMs in clinical settings that need to be addressed, including limited clinical reasoning, variable reliability and accuracy, reproducibility bias, self-serving bias, sponsorship bias, and potential for exacerbating health care disparities. These challenges are further compounded by practical business considerations and infrastructure requirements, including associated costs. To overcome these hurdles and harness the potential of LLMs effectively, this article includes considerations for health care organizations, researchers, and neurologists contemplating the use of LLMs in clinical practice. It is essential for health care organizations to cultivate a culture that welcomes AI solutions and aligns them seamlessly with health care operations. Clear objectives and business plans should guide the selection of AI solutions, ensuring they meet organizational needs and budget considerations. Engaging both clinical and nonclinical stakeholders can help secure necessary resources, foster trust, and ensure the long-term sustainability of AI implementations. Testing, validation, training, and ongoing monitoring are pivotal for successful integration. For neurologists, safeguarding patient data privacy is paramount. Seeking guidance from institutional information technology resources for informed, compliant decisions, and remaining vigilant against biases in LLM outputs are essential practices in responsible and unbiased utilization of AI tools. In research, obtaining institutional review board approval is crucial when dealing with patient data, even if deidentified, to ensure ethical use. Compliance with established guidelines like SPIRIT-AI, MI-CLAIM, and CONSORT-AI is necessary to maintain consistency and mitigate biases in AI research. In summary, the integration of LLMs into clinical neurology offers immense promise while presenting formidable challenges. Awareness of these considerations is vital for harnessing the potential of AI in neurologic care effectively and enhancing patient care quality and safety. The article serves as a guide for health care organizations, researchers, and neurologists navigating this transformative landscape.

Deep learning-assisted diagnosis of benign and malignant parotid tumors based on ultrasound: a retrospective study

BACKGROUND

To develop a deep learning(DL) model utilizing ultrasound images, and evaluate its efficacy in distinguishing between benign and malignant parotid tumors (PTs), as well as its practicality in assisting clinicians with accurate diagnosis.

METHODS

A total of 2211 ultrasound images of 980 pathologically confirmed PTs (Training set: n = 721; Validation set: n = 82; Internal-test set: n = 89; External-test set: n = 88) from 907 patients were retrospectively included in this study. The optimal model was selected and the diagnostic performance evaluation is conducted by utilizing the area under curve (AUC) of the receiver-operating characteristic(ROC) based on five different DL networks constructed at varying depths. Furthermore, a comparison of different seniority radiologists was made in the presence of the optimal auxiliary diagnosis model. Additionally, the diagnostic confusion matrix of the optimal model was calculated, and an analysis and summary of misjudged cases' characteristics were conducted.

RESULTS

The Resnet18 demonstrated superior diagnostic performance, with an AUC value of 0.947, accuracy of 88.5%, sensitivity of 78.2%, and specificity of 92.7% in internal-test set, and with an AUC value of 0.925, accuracy of 89.8%, sensitivity of 83.3%, and specificity of 90.6% in external-test set. The PTs were subjectively assessed twice by six radiologists, both with and without the assisted of the model. With the assisted of the model, both junior and senior radiologists demonstrated enhanced diagnostic performance. In the internal-test set, there was an increase in AUC values by 0.062 and 0.082 for junior radiologists respectively, while senior radiologists experienced an improvement of 0.066 and 0.106 in their respective AUC values.

CONCLUSIONS

The DL model based on ultrasound images demonstrates exceptional capability in distinguishing between benign and malignant PTs, thereby assisting radiologists of varying expertise levels to achieve heightened diagnostic performance, and serve as a noninvasive imaging adjunct diagnostic method for clinical purposes.

Engaging with artificial intelligence in mammography screening: Swedish breast radiologists' views on trust, information and expertise

Objectives

Lack of trust and transparency is stressed as a challenge for clinical implementation of artificial intelligence (AI). In breast cancer screening, AI-supported reading shows promising results but more research is needed on how medical experts, which are facing the integration of AI into their work, reason about trust and information needs. From a sociotechnical information practice perspective, we add to this knowledge by a Swedish case study. This study aims to: (1) clarify Swedish breast radiologists' views on trust, information and expertise pertaining to AI in mammography screening and (2) analytically address ideas about medical professionals' critical engagement with AI and motivations for trust in AI.

Method

An online survey was distributed to Swedish breast radiologists. Survey responses were analysed by descriptive statistical method, correlation analysis and qualitative content analysis. The results were used as foundation for analysing trust and information as parts of critical engagements with AI.

Results

Of the Swedish breast radiologists (n = 105), 47 answered the survey (response rate = 44.8%). 53.2% (n = 25) of the respondents would to a high/somewhat high degree trust AI assessments. To a great extent, additional information would support the respondents' trust evaluations. What type of critical engagement medical professionals are expected to perform on AI as decision support remains unclear.

Conclusions

There is a demand for enhanced information, explainability and transparency of AI-supported mammography. Further discussion and agreement are needed considering what the desired goals for trust in AI should be and how it relates to medical professionals' critical evaluation of AI-made claims in medical decision support.