HyperSORT

Code for the MICCAI 2025 article "HyperSORT: Self-Organising Robust Training with hyper-networks"

Overview

HyperSORT leverages hyper-networks to learn a meaningful latent space parameterizing and discovering the different annotation biases within a training set. The hyper-network maps annotation style latent vectors from the latent space to UNet parameters performing the segmentation task according to the corresponding annotation bias.

TotalSegmentator experiment

The obtained latent space for the TS experiment can be studied here.

Guidelines

1. Create Python environment

Create a Python virtual environment and install the requirements.

Please follow the Pytorch installation guidelines for your specific setting.

Please set-up the free ImFusion license key as specified in the ImFusion python SDK documentation.

2. Download dataset and create data file

In HyperSORT, experiments were conducted on the AMOS dataset and on the TotalSegmentator CT dataset (the corrected V2 version is also available here). Yet, HyperSORT can be used on any dataset. Feel free to apply it to any dataset you are interested in.

Once the training dataset is downloaded, create a data file. A data file is a text file where each line is a tab-separated list of files for a training case.

Templates for such data files can be found in data_files/data_list_TS.txt and data_files/data_list_AMOS.txt.

3. Create a training yaml config

Create a training yaml config parameterizing your experiments. The used configs for our experiments can be found in: configs/AMOS.yaml and configs/TS.yaml.

A config yaml file must specify the following sections:

OutputPath: Specifying the folder where the experiment training plots and results will be saved.

DataFile: The path to the created data file specifying the files constituting the training set to learn the annotation style latent space from.

Pipeline: A list of preprocessing operations to be applied to the image. We used the ImFusion SDK to preprocess images. Documentation on available operations can be found here. Each preprocessing operation is specified as:

    - OperationName: # without the suffix "Operation"
        OperationParameterName1: OperationParameterValue1
        ... 

ModelConfig: Specify HyperSORT's model parameters

• HyperNetworkLayers is a list starting with the latent space dimension followed by the hyper-network hidden layers dimensions

• The UNet section specifies the UNet archtiecture's parameters

4. Launch HyperSORT training

Run:

python main.py -c path_to_config_file.yaml