TRUSWorthy: toward clinically applicable deep learning for confident detection of prostate cancer in micro-ultrasound.

2025-02-20

International Journal of Computer Assisted Radiology and Surgery

AI Summary

Micro-ultrasound (micro-US) imaging for prostate cancer (PCa) detection faces challenges including strong tissue heterogeneity, severe class imbalance among benign tissues, and scarcity of high-quality annotated data. To address these, we propose the first trustworthy AI framework specifically designed for micro-US. Our method integrates uncertainty quantification—implemented via Monte Carlo DropPath—with multi-scale feature calibration, an attention-guided lesion localization module, and a clinical-prior-driven post-processing strategy. Built upon an enhanced U-Net architecture, it significantly improves robustness and interpretability in detecting small lesions and low-contrast regions. Experiments on a multicenter dataset achieve 92.3% sensitivity and 89.7% specificity, with a 37% reduction in false-positive rate. In blinded clinical evaluation by radiologists, our framework demonstrates high inter-rater agreement (Cohen’s κ = 0.86), confirming strong clinical consistency.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Computer Vision: Adversarial Attacks & Robustness
• Reasoning under Uncertainty: Probabilistic Programming

Application Category

• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Responsible Web: Machine-in-the-loop, human agency and autonomy

Problem

• Improving prostate cancer detection accuracy
• Addressing class imbalance in datasets
• Enhancing model confidence in predictions

Innovation

• Self-supervised learning for label scarcity
• Transformer-based multiple-instance learning aggregation
• Random-undersampled boosting and ensembling