Michael Uder

Sebastian Wind, Tri-Thien Nguyen, Jeta Sopa et al.

Clinical LLMs are often scaled by increasing model size, context length, retrieval complexity, or inference-time compute, with the implicit expectation that higher accuracy implies safer behavior. This assumption is incomplete in medicine, where a few confident, high-risk, or evidence-contradicting errors can matter more than average benchmark performance. We introduce SaFE-Scale, a framework for measuring how clinical LLM safety changes across model scale, evidence quality, retrieval strategy, context exposure, and inference-time compute. To instantiate this framework, we introduce RadSaFE-200, a Radiology Safety-Focused Evaluation benchmark of 200 multiple-choice questions with clinician-defined clean evidence, conflict evidence, and option-level labels for high-risk error, unsafe answer, and evidence contradiction. We evaluated 34 locally deployed LLMs across six deployment conditions: closed-book prompting (zero-shot), clean evidence, conflict evidence, standard RAG, agentic RAG, and max-context prompting. Clean evidence produced the strongest improvement, increasing mean accuracy from 73.5% to 94.1%, while reducing high-risk error from 12.0% to 2.6%, contradiction from 12.7% to 2.3%, and dangerous overconfidence from 8.0% to 1.6%. Standard RAG and agentic RAG did not reproduce this safety profile: agentic RAG improved accuracy over standard RAG and reduced contradiction, but high-risk error and dangerous overconfidence remained elevated. Max-context prompting increased latency without closing the safety gap, and additional inference-time compute produced only limited gains. Worst-case analysis showed that clinically consequential errors concentrated in a small subset of questions. Clinical LLM safety is therefore not a passive consequence of scaling, but a deployment property shaped by evidence quality, retrieval design, context construction, and collective failure behavior.

Kubilay Can Demir, Matthias May, Axel Schmid et al.

This paper presents a new multimodal interventional radiology dataset, called PoCaP (Port Catheter Placement) Corpus. This corpus consists of speech and audio signals in German, X-ray images, and system commands collected from 31 PoCaP interventions by six surgeons with average duration of 81.4 $\pm$ 41.0 minutes. The corpus aims to provide a resource for developing a smart speech assistant in operating rooms. In particular, it may be used to develop a speech controlled system that enables surgeons to control the operation parameters such as C-arm movements and table positions. In order to record the dataset, we acquired consent by the institutional review board and workers council in the University Hospital Erlangen and by the patients for data privacy. We describe the recording set-up, data structure, workflow and preprocessing steps, and report the first PoCaP Corpus speech recognition analysis results with 11.52 $\%$ word error rate using pretrained models. The findings suggest that the data has the potential to build a robust command recognition system and will allow the development of a novel intervention support systems using speech and image processing in the medical domain.

Tri-Thien Nguyen, Lorenz A. Kapsner, Tobias Hepp et al.

Background: Magnetic resonance imaging (MRI) has high sensitivity for breast cancer detection, but interpretation is time-consuming. Artificial intelligence may aid in pre-screening. Purpose: To evaluate the DINOv2-based Medical Slice Transformer (MST) for ruling out significant findings (Breast Imaging Reporting and Data System [BI-RADS] >=4) in contrast-enhanced and non-contrast-enhanced abbreviated breast MRI. Materials and Methods: This institutional review board approved retrospective study included 1,847 single-breast MRI examinations (377 BI-RADS >=4) from an in-house dataset and 924 from an external validation dataset (Duke). Four abbreviated protocols were tested: T1-weighted early subtraction (T1sub), diffusion-weighted imaging with b=1500 s/mm2 (DWI1500), DWI1500+T2-weighted (T2w), and T1sub+T2w. Performance was assessed at 90%, 95%, and 97.5% sensitivity using five-fold cross-validation and area under the receiver operating characteristic curve (AUC) analysis. AUC differences were compared with the DeLong test. False negatives were characterized, and attention maps of true positives were rated in the external dataset. Results: A total of 1,448 female patients (mean age, 49 +/- 12 years) were included. T1sub+T2w achieved an AUC of 0.77 +/- 0.04; DWI1500+T2w, 0.74 +/- 0.04 (p=0.15). At 97.5% sensitivity, T1sub+T2w had the highest specificity (19% +/- 7%), followed by DWI1500+T2w (17% +/- 11%). Missed lesions had a mean diameter <10 mm at 95% and 97.5% thresholds for both T1sub and DWI1500, predominantly non-mass enhancements. External validation yielded an AUC of 0.77, with 88% of attention maps rated good or moderate. Conclusion: At 97.5% sensitivity, the MST framework correctly triaged cases without BI-RADS >=4, achieving 19% specificity for contrast-enhanced and 17% for non-contrast-enhanced MRI. Further research is warranted before clinical implementation.

Yixing Huang, Christoph Bert, Philipp Sommer et al.

Brain metastases occur frequently in patients with metastatic cancer. Early and accurate detection of brain metastases is very essential for treatment planning and prognosis in radiation therapy. To improve brain metastasis detection performance with deep learning, a custom detection loss called volume-level sensitivity-specificity (VSS) is proposed, which rates individual metastasis detection sensitivity and specificity in (sub-)volume levels. As sensitivity and precision are always a trade-off in a metastasis level, either a high sensitivity or a high precision can be achieved by adjusting the weights in the VSS loss without decline in dice score coefficient for segmented metastases. To reduce metastasis-like structures being detected as false positive metastases, a temporal prior volume is proposed as an additional input of DeepMedic. The modified network is called DeepMedic+ for distinction. Our proposed VSS loss improves the sensitivity of brain metastasis detection for DeepMedic, increasing the sensitivity from 85.3% to 97.5%. Alternatively, it improves the precision from 69.1% to 98.7%. Comparing DeepMedic+ with DeepMedic with the same VSS loss, 44.4% of the false positive metastases are reduced in the high sensitivity model and the precision reaches 99.6% for the high specificity model. The mean dice coefficient for all metastases is about 0.81. With the ensemble of the high sensitivity and high specificity models, on average only 1.5 false positive metastases per patient needs further check, while the majority of true positive metastases are confirmed. The ensemble learning is able to distinguish high confidence true positive metastases from metastases candidates that require special expert review or further follow-up, being particularly well-fit to the requirements of expert support in real clinical practice.