ORACLE-CT: Anatomy-Aware Support Pooling for CT Classification
Researchers have developed ORACLE-CT, a novel framework designed to enhance the accuracy of classifying diseases from abdominal CT scans. This system leverages multi-organ segmentation to guide attention pooling towards relevant anatomical regions, addressing the challenge of localized evidence within large 3D volumes. Evaluations showed that ORACLE-CT, when integrated with various encoders like DINOv3 and I3D-ResNet-121, significantly improved classification performance and external robustness compared to standard global pooling methods. AI
IMPACT Enhances diagnostic accuracy in medical imaging by focusing AI on relevant anatomical evidence.