Why is patch size important? Dealing with context variability in segmentation networks

Marc Balle Sanchez, Samuel JOUTARD, Mohammad Farid Azampour, Raphael Prevost

Keywords: Hypernetworks, patch size augmentation

TL;DR: We demonstrate that patch size augmentation and hypernetworks conditioning the model on the patch size are both effective solutions against distribution shift in contextual information when performing inference with diverse patch sizes.

Abstract:

Deep learning models have made significant advancements in medical image segmentation. Patch-based training is the standard practice for 3D segmentation models. During model deployment in hospital settings, resource constraints may require performing inference with
reduced patch sizes. However, this might lead to a decrease in performance. In this study we demonstrate that patch size augmentation is a straightforward and effective approach to enhance the robustness of a 3D U-Net to different patch sizes during inference. Furthermore,
we show that using a hypernetwork to adapt the U-Net to diverse patch sizes further enhances performance across the patch size spectrum.

Submission Number: 109