This repository contains the official implementation of our paper: CSAL-3D: Cold-start Active Learning for 3D Medical Image Segmentation via SSL-driven Uncertainty-Reinforced Diversity Sampling, for 28th International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI 2025, Early Accept).
The overall CSAL-3D pipeline consists of:
- A CSAL-adapted Self-Supervised Learning (SSL) framework for both 3D-aware feature extraction and uncertainty estimation.
- An Ensemble-based Uncertainty Estimation strategy to generate sample-level uncertainty scores.
- A URDS (Uncertainty-Reinforced Diversity Sampling) method that hierarchically combines diversity and uncertainty for one-shot sample selection.
We evaluate our method on three publicly available 3D medical image segmentation datasets from the Medical Segmentation Decathlon (MSD):
- Brain Tumor (Task01_BrainTumour) [MRI, multi-modality]
- Heart (Task02_Heart) [MRI]
- Spleen (Task09_Spleen) [CT]
- Datasets can be downloaded from the official MSD website: 👉 http://medicaldecathlon.com/
We recommend using Python 3.8+ with PyTorch 2.3.1 and MONAI 1.3+.
Our codebase is built upon the excellent COLosSAL project (https://github.com/han-liu/COLosSAL), which provides a benchmark for Cold-Start Active Learning in 3D medical image segmentation.