MrGiovanni/SuPreM

What it solves

SuPreM provides a suite of pre-trained 3D models and a large-scale annotated dataset (AbdomenAtlas 1.1) to improve the performance and efficiency of 3D medical imaging tasks. It addresses the difficulty of obtaining large, detailed-annotated datasets for supervised pre-training in medical AI, which is often more effective than self-supervised alternatives.

How it works

The project consists of two primary components:

1. AbdomenAtlas 1.1: A dataset containing 9,262 CT volumes with per-voxel annotations for 25 organs and pseudo annotations for seven types of tumors.
2. SuPreM Models: A collection of pre-trained weights for several widely recognized 3D AI model backbones, including Swin UNETR, U-Net, and SegResNet, trained on the AbdomenAtlas dataset.

Users can use these pre-trained weights to perform direct inference on unseen CT volumes or fine-tune the models for specific downstream medical tasks, such as organ, muscle, vertebrae, cardiac, and rib segmentation or pancreas tumor detection.

Who it’s for

Medical imaging researchers and AI developers working with 3D volumetric medical image analysis, specifically those focusing on organ and tumor segmentation in CT scans.

Highlights

• Large-Scale Supervised Pre-training: Utilizes a Been-annotated dataset of over 9,000 CT volumes to provide a foundation for 3D medical segmentation.
• Multiple Backbone Support: Provides pre-trained weights for Swin UNETR, U-Net, and SegResNet.
• Broad Transferability: Demonstrates effectiveness across various downstream tasks including multi-organ and tumor segmentation.
• Integrated Dataset: Includes the AbdomenAtlas 1.1 dataset with detailed per-voxel annotations for 25 organs.