GammaNet Code for Cardiac Images Segmentation

Usage

• The whole architecture is config driven. Define the configuration for your experiment in ./config directory. See default.json to see the structure and available options.
• After defining the configuration (say my_exp.json) - simply run python3 main.py my_exp
• If you want to run an experiment with a different name than the one mentioned in config - run python3 main.py my_exp instance_name. This is helpful when you wish to do multiple runs for the same experiment.
• This will start the experiment. It will also take care of building the dataset, if it doesn't exist already in mentioned path.
• The logs, stats, plots and saved models would be stored in ../experiment_data/my_exp dir.
• To resume an ongoing experiment, simply run the same command again. It will load the latest stats and models and resume training.

Config Options

This is a sample config.

{
  "experiment_name": "default_experiment",
  "dataset": {
    "training_fraction": 0.8,
    "path": "../processed_data/regular",
    "timeseries": false,
    "transforms": [
      "GaussianSmooth",
      "CLAHE",
      "MinMax",
      "PadOrCenterCrop"
    ]
  },
  "experiment": {
    "num_epochs": 200,
    "learning_rate": 1e-3,
    "batch_size_train": 2,
    "batch_size_val": 4,
    "num_workers": 8
  },
  "model": {
    "type": "GammaNet",
    "criterion": "bce",
    "timeseries": false,
    "ensemble": false,
    "fgru_timesteps": 4
  }
}

Numeric and Boolean attributes can be varied as expected. Possible options for configurable attributes-

• model.type: GammaNet or UNet
• model.criterion: bce, dice, weighted, combined

Supported Datasets

• ACDC

Files

• main.py: Main driver class
• experiment.py: Main experiment class. Initialized based on config - takes care of training, saving stats and plots, logging and resuming experiments.
• gammanet.py: main gamma-net implementation
• fgru.py: implementation of fGRU
• dataset.py: simple Dataset class for PyTorch
• dataset_factory: Factory to build datasets based on config
• dataset_maker.py: conversion scripts to process different datasets into a format readable by the SimpleDataset class
• transforms.py: image preprocessing transforms
• transforms_factory.py: Factory to build image transforms based on config
• criteria.py: scoring functions
• constants.py: constants used across the project
• file_utils.py: utility functions for handling files

---

Original paper:

Linsley, Drew, Junkyung Kim, and Thomas Serre. "Sample-efficient image segmentation through recurrence." arXiv preprint arXiv:1811.11356 (2018).