Using Swin UNETR as a Classifier: Why the Segmentation Architecture Works for Weak Supervision

[Full post content coming soon. This post will cover the architectural choice behind the ABCD pipeline: why CNNs fail on white matter lesions that span distant regions, how shifted-window attention provides linear-complexity long-range dependencies, and how Global Average Pooling converts a segmentation backbone into a classification head that produces Grad-CAM-compatible feature maps.]