Notes

April 2026Weak Supervision for Medical Imaging: From Scan-Level Labels to Voxel-Level Maps

When you have 11,868 brain scans and zero voxel annotations, you need a different playbook. Here is how Global Average Pooling, Grad-CAM, and pseudo-masks chain together into a working pipeline.

March 2026DBS Connectivity Fingerprints: What the Data Actually Says

An honest account of testing the destination network convergence hypothesis on 116 stimulation configurations across 5 patients — and why the answer is more interesting than a clean positive result.

February 2026VICReg in the Wild: Lessons from Applying Distribution Regularization to fMRI

The most interesting result from the Brain-JEPA extension was not the AUC improvement — it was the variance reduction. Here is what that means and why it matters.

January 2026A Practical Guide to the ABCD Dataset for ML Researchers

The ABCD Study has 11,868 participants, 239 tabulated data tables, and 92,419 variables. Here is how to navigate it without losing your mind.

December 2025Using Swin UNETR as a Classifier: Why the Segmentation Architecture Works for Weak Supervision

Swin UNETR was designed for segmentation. Here is the argument for why its hierarchical transformer encoder is exactly what you want for weakly-supervised 3D medical image classification.