Multichannel Model Support for Pre-Aligned Volumes - #1896

Overview

Adds ability for training multichannel segmentation models on pre-aligned/co-registered volumes using the local datastore. I was working based off this issue #241

I added a working implementation for the BraTS 2020 dataset with individual volumes as a reference use case.

Changes

• Added optional --input_channels argument to support multichannel model configurations
• Added optional --multi_file argument to handle multi-file volume inputs (each sample has a directory with multiple volumes)
• Added optional --multichannel argument to handle multichannel inputs (example sample has a single image with multiple volumes)

Results (BraTS 2020 — 150 epochs)

Notes

• Tested on BraTS 2020 with 4-channel input (T1, T1ce, T2, FLAIR)
• Volumes must be pre-aligned/co-registered before use

I'd be happy to incorporate any feedback from maintainers so this could be added

Summary by CodeRabbit

• New Features

  • Multi-channel (4D) and multi-file (directory-based) dataset support across app, CLI, datastore, plugin, and training/inference pipelines.
  • New BraTS multilabel brain tumor segmentation task: training, inference, transforms, loaders, and utilities.
  • Local datastore: NRRD support and ability to add directories/files.

• Behavior Changes

  • Flags (multichannel, multi_file, input_channels) propagated through app, CLI, training, inference, and active learning; upload endpoint blocks multi-file uploads.

• Documentation

  • Added BraTS usage block to sample app README (duplicated).