How adigasu/anatomically-aware_uncertainty_for_semi-supervised_segmentation works
Project type and purpose
• This is a research‐oriented, semi-supervised medical image segmentation codebase.
• It implements a two-stage pipeline:
a) Train a denoising autoencoder (DAE) to learn anatomical representations.
b) Train a mean-teacher segmentation model that leverages the DAE embeddings to estimate uncertainty and improve performance in a low-label regime.
File structure
• Top-level:
• LICENSE, README.md, requirement.txt
• code_DAE/ – DAE training and tests
• code_DAE_MT/ – Mean-teacher segmentation with DAE embeddings
• Inside each of code_DAE & code_DAE_MT:
• dataloaders/ – data loading logic (medical_dataloader.py + helper utils.py)
• networks/ – model definitions (vnet.py for 3D V-Net architecture)
• utils/ – losses.py, ramps.py, util.py
• train_*.py – training scripts
• test_*.py – evaluation scripts
• Requirements imply PyTorch, tensorboardX, numpy, skimage, etc.
README for additional insights
• Two phases of training are clearly described, with example CLI commands.
• References to datasets (LA 2018, FLARE 2021).
Additional notes
• External data sources as “outside” elements feeding into the Data Loader.