Weak Supervision for Medical Imaging: From Scan-Level Labels to Voxel-Level Maps

April 2026

[Full post content coming soon. This post will cover the weak supervision strategy used in the ABCD leukoaraiosis pipeline: why APLoss outperforms cross-entropy under extreme class imbalance, how Grad-CAM 3D saliency maps localize lesions without masks, and what Otsu thresholding buys you when used to generate pseudo-labels for self-training.]

← Notes