NeurIPS 2025 Submission

This repository contains the implementation code for our NeurIPS 2025 submission.

🔒 Anonymous Review Notice

This repository is anonymized for double-blind review.
Author information and license will be added upon acceptance.
Full code and training scripts will be released after the paper is accepted.

Description

We propose a novel framework for semi-supervised medical image segmentation using chained diffusion-based pseudo-label refinement.

Requirements

• Python 3.8+
• PyTorch >= 1.10
• CUDA 11.x
• Other dependencies are listed in requirements.txt (to be provided if necessary)

Data Preparation

• ACDC: Download from the official SSL4MIS repository: https://github.com/HiLab-git/SSL4MIS/tree/master/data/ACDC

• MS-CMRSEG19: Download from the official challenge website: https://zmiclab.github.io/zxh/0/mscmrseg19/

• Task05 Prostate (Medical Segmentation Decathlon): http://medicaldecathlon.com/

Usage

Train the model with the following command:

Example: Train the segmentation model on ACDC with 1 labeled patient

CUDA_VISIBLE_DEVICES=0 python train_diffrect_chain_ACDC.py
--exp ACDC/diffrect
--labelnum 1
--num_classes 4
--root_path ./datasets/ACDC

Acknowledgement

We sincerely appreciate the following open-source projects for their valuable contributions, which our work builds upon:

• DiffRect
• U-KAN
• SSL4MIS
• guided-diffusion
• GSS
• DiffUNet