Why is patch size important?

Code for the MIDL 2025 paper "Why is patch size important? Dealing with context variability in segmentation networks"

Overview

In this repository, we propose two solutions that address the distribution shift in contextual information that arises when using different patch sizes at inference time compared to training. Using a different patch size, specially a smaller one, is a practical stategy to decrease computational load when deploying a model in resource-contrained environments like hospitals. The two proposed solutions are:

1. Patch size augmentation during training—a simple yet effective approach to improve robustness.

2. A hypernetwork conditioned on the patch size, which dynamically generates optimal weights for the primary model based on the available context. This method provides additional performance gains, with a modest increase in training complexity.

Repository organization

This repository is composed of 5 files:

• data.py: This file contains two data loaders, one for training (patched images) and another one for inference (whole images). This part of the pipeline relies on the ImFusion Python SDK for image loading and basic preprocessing. These loaders expect data list files that specify input data files and the associated segmentation file names. Example data lists are available in the data_files folder.

• models.py: This file contains the model definitions, in particular the definition of a standard UNet and a conditional UNet which are the two types of architectures used in our experiments in different training or inference settings.

• train.py: This file contains the training and validation loops. To reproduce our training, simply run:

  python train.py -c configs/config.yaml

  Notice that the sample data files in data_files folder must be updated with the correct file paths.

• eval.py: This file contains our evaluation script which we used to compute the metrics reported in the paper.

• utils.py: This file contains various utility functions including the losses, image saving functionalities (relying on the ImFusion Python SDK as well), and others.