Webpages tagged with «tmlr»

Overview of memory-constrained full-bag E2E training for WSI MIL. Frozen training blocks encoder gradients, while naive E2E training retains full-bag encoder activations and can exceed GPU memory. Our CNN path partitions memory-intensive layers and offloads state to CPU RAM; our ViT/foundation model path computes embeddings in micro-batches and recomputes the encoder during backpropagation. Both paths update the MIL head and all trainable encoder layers using all selected tiles.

TMLR Paper: "To Freeze or Not to Freeze? Memory-Constrained End-to-End Training for Whole-Slide Image Classification in Histopathology"

Published Sep. 30, 2026 11:41 PM

Deep-learning models work well on medical images, but whole-slide images (WSIs) used in histopathology are so large that researchers usually have to freeze much of the model (the whole encoder) during training. This work proposes a novel method to train the full model end-to-end under realistic GPU memory constraints and investigates when this is beneficial.