MEDA is a lightweight decision-adaptation framework for multi-target CT recognition. It keeps a shared 3D CT representation fixed and learns target-specific decision rules by selecting both a scoring function and an operating threshold for each diagnostic endpoint.
17 CT targets • 0.738 mean balanced accuracy • 0.746 mean AUROC • 15/17 targets improved by operating-point calibration
| Region | Targets | Validation N | Positives | BalAcc | AUROC |
|---|---|---|---|---|---|
| Abdomen | 15 | 1,634 | 817 | 0.742 | 0.750 |
| Chest | 2 | 264 | 132 | 0.708 | 0.715 |
| Overall | 17 | 1,898 | 949 | 0.738 | 0.746 |
Operating-point calibration improved mean balanced accuracy from 0.635 to 0.738, with gains in 15/17 targets.