Derya Yakar

In this study of 22,445 prostate MRI examinations from 22 countries, an AI system’s performance was comparable to radiology readings for prostate cancer detection in external diagnostic and screening datasets.

Artificial Intelligence (AI) for detecting clinically significant prostate cancer (csPCa) on MRI has achieved diagnostic performance comparable to that of radiologists. By autonomously interpreting examinations, AI could improve workflow efficiency and help address increasing imaging demands and radiologist shortages. Despite this promise, autonomous AI has not been implemented in clinical practice. This narrative review explores remaining technical and societal barriers to deploying autonomous csPCa detection. We focus on three key domains: limitations in the current evidence base, safety issues and mitigation strategies, and the perspectives of patients and radiologists. Our findings highlight the need for evidence from large, multicenter, prospective trials and evaluation frameworks that reflect the consequences of clinical decision-making, as well as further exploration of safeguards to monitor and address mismatches between training data and incoming scans during deployment. Moreover, patients and radiologists show limited acceptance of autonomous AI, although this may improve with greater transparency, targeted education, and clearer guidelines on medico-legal responsibilities. Addressing these challenges is essential to the responsible deployment of autonomous AI and to realizing its efficiency gains in clinical practice.

Background: Manual lesion measurements remain the standard for assessing oncologic treatment response, despite being time-consuming and prone to substantial inter-reader variability, which may lead to inconsistent response classification and subsequent variability in treatment decisions. Purpose: To evaluate the impact of an artificial intelligence (AI) system for assisted lesion measurement on reading time and measurement consistency in follow-up CT examinations using the Response Evaluation Criteria in Solid Tumors (RECIST 1.1). Methods and Materials: In this retrospective reader study, follow-up chest-abdomen-pelvis CT examinations from 212 oncology patients collected at two Dutch hospitals were assessed by 23 readers (15 radiologists, 8 residents) recruited from 11 international institutions under three conditions: unassisted, AI-assisted, and expert-assisted (using a prior radiologist’s unassisted measurement). To prevent bias related to the source of the measurements, readers were informed that all support was AI-derived. Primary outcomes were reading time to completion and inter-observer measurement variability. For each outcome, a Bayesian generalized linear mixed model was used to analyze the results. Results: AI assistance significantly reduced per-patient reading time versus unassisted reading (-36.0 s; 95% CI: -53.0, -22.1). At the lesion level, it was associated with a small increase in variability relative to the expert-derived reference standard (1.32 mm; 95% CI: 0.83, 1.91). At the patient level, AI assistance did not meaningfully affect change in the sum of longest diameters (SLD; -0.38 mm; 95% CI: -1.57, 0.72), but increased RECIST outcome agreement by 7.7% (95% CI: 2.8, 12.7) compared with unassisted reading. Expert-assisted reading yielded even higher inter-reader agreement (13.3%; 95% CI: 8.6, 18.1). Conclusion: AI assistance reduced reading time and improved patient-level RECIST agreement, despite a small increase in lesion-level measurement variability. These findings suggest that AI-assisted RECIST assessment may improve workflow and response classification consistency, while also providing a benchmark from expert-assisted reading for future AI development.

Purpose: To clarify overlapping post-market obligations under the EU Artificial Intelligence Act (AIA) and EU Medical Device Regulation (MDR) for high-risk artificial intelligence (AI) Software as a Medical Device (SaMD), and to map the regulatory landscape for manufacturers, healthcare providers, AI providers, and AI deployers. Methods: We conducted a qualitative doctrinal legal analysis of post-market provisions in the AIA and MDR, using a case study of a high-risk Class III AI SaMD for prostate cancer radiology. No empirical clinical or performance data were collected. The analysis focused on key stakeholders, including device manufacturers and deployers (e.g., healthcare providers). We sought to identify (1) convergence, where both regulations impose overlapping or complementary requirements, and (2) divergence, where obligations are addressed by only one regulation, revealing potential regulatory gaps. Results: We organized the extracted post-market obligations into ten categories. Overall, both regulations place increasing emphasis on lifecycle traceability and continuous monitoring. We identified convergence in areas such as documentation and performance monitoring, while divergences emerged in domains like human oversight (in the AIA) and reporting non-serious patterns (in the MDR). We also identified gaps in regulatory guidance, particularly regarding system updates, human oversight, and the evolving responsibilities of healthcare providers. Conclusion: The AIA and MDR share common ground in some post-market areas but also diverge in key responsibilities. To ensure safe and effective use of high-risk AI in healthcare, clearer coordination between the two frameworks is needed, especially in areas such as human oversight and system modification, where current guidance remains limited.

Federated learning enables collaborative model training, enhancing the performance and generalizability of trained models for MRI-based prostate gland segmentation and clinically significant prostate cancer detection compared with locally trained models.

Background: Computed tomography urography (CTU) is routinely used to evaluate the upper urinary tract in patients with hematuria. CTU may detect incidental findings outside the urinary tract, but it remains unclear if this adds value. This study aimed to develop a deep learning algorithm that automatically segments and selectively visualizes the urinary tract on CTU. Methods: The urinary tract (kidneys, ureters, and urinary bladder) was manually segmented on 2 mm dual-phase CTU slices of 111 subjects. With this dataset, a deep learning-based AI was trained to automatically segment and selectively visualize the urinary tract on CTU scans (including accompanying unenhanced CT scans), which we dub “focused view CTU”. Focused view CTU was technically optimized and tested in 39 subjects with hematuria. Results: The technically optimized focused view CTU algorithm provided complete visualization of 97.4% of kidneys, 80.8% of ureters, and 94.9% of urinary bladders. All urinary tract organs were completely visualized in 66.6% of cases. In these cases (excluding 33.3% of cases with incomplete visualization), focused view CTU intrinsically achieved a sensitivity, specificity, positive predictive value, and negative predictive value of 100.0%, 92.3%, 92.9%, and 100.0% for lesions in the urinary tract compared to unmodified CT, although interrater agreement was moderate (κ = 0.528). All incidental findings were successfully hidden by focused view CTU. Conclusions: Focused view CTU provides adequate urinary tract segmentation in most cases, but further research is needed to optimize the technique (segmentation does not succeed in about one-third of cases). It offers selective urinary tract visualization, potentially aiding in assessing relevance and cost-effectiveness of detecting incidental findings in hematuria patients through a prospective randomized trial.

Background/Objectives: There is a growing need for efficient prostate MRI protocols due to their increasing use in managing prostate cancer (PCa) and potential inclusion in screening. Deep learning reconstruction (DLR) may enhance MR acquisitions and improve image quality compared to conventional acceleration techniques. This systematic review examines DLR approaches to prostate MRI. Methods: A search of PubMed, Web of Science, and Google Scholar identified eligible studies comparing DLR to conventional reconstruction for prostate imaging. A narrative synthesis was performed to summarize the impact of DLR on acquisition time, image quality, and diagnostic performance. Results: Thirty-three studies showed that DLR can reduce acquisition times for T₂w and DWI imaging while maintaining or improving image quality. It did not significantly affect clinical tasks, such as biopsy decisions, and performed comparably to human readers in PI-RADS scoring and the detection of extraprostatic extension. However, AI models trained on conventional data might be less accurate with DLR images. The heterogeneity in image quality metrics among the studies prevented quantitative synthesis. Discussion: DLR has the potential to achieve substantial time savings in prostate MRI while maintaining image quality, which is especially relevant because of increased MRI demands. Future research should address the effect of DLR on clinically relevant downstream tasks, including AI algorithms’ performances and biopsy decisions, and explore task-specific accelerated protocols for screening, image-guided biopsy, and treatment.

PurposeInterventional MR imaging struggles with speed and efficiency. We aim to accelerate transrectal in-bore MR-guided biopsies for prostate cancer through undersampled image reconstruction and instrument localization by image segmentation.ApproachIn this single-center retrospective study, we used 8464 MR 2D multislice scans from 1289 patients undergoing a prostate biopsy to train and test a deep learning–based spatiotemporal MR image reconstruction model and a nnU-Net segmentation model. The dataset was synthetically undersampled using various undersampling rates (R=8, 16, 25, 32). An annotated, unseen subset of these data was used to compare our model with a nontemporal model and readers in a reader study involving seven radiologists from three centers based in the Netherlands. We assessed a maximum noninferior undersampling rate using instrument prediction success rate and instrument tip position (ITP) error.ResultsThe maximum noninferior undersampling rate is 16-times for the temporal model (ITP error: 2.28 mm, 95% CI: 1.68 to 3.31, mean difference from reference standard: 0.63 mm, P=.09), whereas a nontemporal model could not produce noninferior image reconstructions comparable to our reference standard. Furthermore, the nontemporal model (ITP error: 6.27 mm, 95% CI: 3.90 to 9.07) and readers (ITP error: 6.87 mm, 95% CI: 6.38 to 7.40) had low instrument prediction success rates (46% and 60%, respectively) compared with the temporal model’s 95%.ConclusionDeep learning–based spatiotemporal MR image reconstruction can improve time-critical intervention tasks such as instrument tracking. We found 16 times undersampling as the maximum noninferior acceleration where image quality is preserved, ITP error is minimized, and the instrument prediction success rate is maximized.

The proposed deep learning model based on MRI and clinical characteristics showed good performance in predicting progression from low-risk prostate cancer to clinically significant prostate cancer.