Bhat S, Maier A (2026)
Publication Language: English
Publication Type: Conference contribution, Conference Contribution
Publication year: 2026
Publisher: Springer
Conference Proceedings Title: 28th International Conference of Medical Image Computing and Computer Assisted Intervention – MICCAI
Event location: Daejeon, South Korea
Computer-aided diagnosis (CAD) systems play a pivotal role in medical imaging, particularly in radiographic research. This work focusses on practical improvements to CAD systems by addressing three key challenges. These include 1) enhancing sensitivity at high specificities, 2) harnessing ad 3) improving precise lesion detection. First, we propose an AUC-reshaping algorithm that enhances sensitivity at high specificity thresholds, aligning to the operating conditions of CAD systems in real-world clinical settings. Second, we introduce Patch-CLIP, for coarse localization using classification networks, via Vision-Language (VL) self-supervised learning. For zero-shot classification, we introduce CXR-CML to explicitly cluster the VL latent space and improve results for long-tailed classes. Finally, we present EM-DETR, enabling robust, precise lesion detection by learning class-specific “exemplar” features. Our contributions demonstrate improvements in diagnostic accuracy, classification, localization, and lesion detection. In our experiments with CXR, mammography, and credit card fraud classification, AUCReshaping increases sensitivity at 98% specificity by 2–9% points; Patch-CLIP improves classification by 1% point and sensitivity at 0.5 False Positives (FP) by 3-7% points on public and private CXR datasets; CXR-CML yields a 7% point increase in zero-shot classification; and EM-DETR yields precision gains of 4–16% points across CXR, mammograms, and stenosis datasets with a 2x increase in True Positive (TP) on an out-ofdistribution (OOD) mammogram dataset. Together, these methods address key limitations of current CAD systems, enhancing their reliability and clinical applicability.
APA:
Bhat, S., & Maier, A. (2026). Improving Clinical Usability of Deep Learning for Radiographic Diagnosis: Learning under Constraints of Specificity, Data, and Annotation. In 28th International Conference of Medical Image Computing and Computer Assisted Intervention – MICCAI. Daejeon, South Korea: Springer.
MLA:
Bhat, Sheethal, and Andreas Maier. "Improving Clinical Usability of Deep Learning for Radiographic Diagnosis: Learning under Constraints of Specificity, Data, and Annotation." Proceedings of the MICCAI 2025, Daejeon, South Korea Springer, 2026.
BibTeX: Download