brilliant-b / suprem_exp Goto Github PK

We developed a suite of pre-trained 3D models, named SuPreM, that combined the best of large-scale datasets and per-voxel annotations, showing the transferability across a range of 3D medical imaging tasks.

How Well Do Supervised 3D Models Transfer to Medical Imaging Tasks?
Wenxuan Li, Alan Yuille, and Zongwei Zhou^*
Johns Hopkins University
International Conference on Learning Representations (ICLR) 2024 (oral; top 1.2%)
paper | code | slides | talk

Transitioning to Fully-Supervised Pre-Training with Large-Scale Radiology ImageNet for Improved AI Transferability in Three-Dimensional Medical Segmentation
Wenxuan Li¹, Junfei Xiao¹, Jie Liu², Yucheng Tang³, Alan Yuille¹, and Zongwei Zhou^1,*
1Johns Hopkins University
²City University of Hong Kong
³NVIDIA
Radiological Society of North America (RSNA) 2023
abstract | code | slides | talk

★ We have maintained a document for Frequently Asked Questions.

★ We have reviewed 3D medical pre-training in Awesome Medical Pre-Training.

The release of AbdomenAtlas 1.0 can be found at https://github.com/MrGiovanni/AbdomenAtlas

AbdomenAtlas 1.1 is an extensive dataset of 9,262 CT volumes with per-voxel annotation of 25 organs and pseudo annotations for seven types of tumors, enabling us to finally perform supervised pre-training of AI models at scale. Based on AbdomenAtlas 1.1, we also provide a suite of pre-trained models comprising several widely recognized AI models.

Prelimianry benchmark showed that supervised pre-training strikes as a preferred choice in terms of performance and efficiency compared with self-supervised pre-training.

We anticipate that the release of large, annotated datasets (AbdomenAtlas 1.1) and the suite of pre-trained models (SuPreM) will bolster collaborative endeavors in establishing Foundation Datasets and Foundation Models for the broader applications of 3D volumetric medical image analysis.

The following is a list of supported model backbones in our collection. Select the appropriate family of backbones and click to expand the table, download a specific backbone and its pre-trained weights (name and download), and save the weights into ./pretrained_weights/. More backbones will be added along time. Please suggest the backbone in this channel if you want us to pre-train it on AbdomenAtlas 1.1 containing 9,262 annotated CT volumes.

Swin UNETR

name	params	pre-trained data	resources	download
Tang et al.	62.19M	5050 CT		weights
Jose Valanaras et al.	62.19M	50000 CT/MRI		weights
Universal Model	62.19M	2100 CT		weights
SuPreM	62.19M	2100 CT	ours 🌟	weights

U-Net

name	params	pre-trained data	resources	download
Models Genesis	19.08M	623 CT		weights
UniMiSS	tiny	5022 CT&MRI		weights
	small	5022 CT&MRI		weights
Med3D	85.75M	1638 CT		weights
DoDNet	17.29M	920 CT		weights
Universal Model	19.08M	2100 CT		weights
SuPreM	19.08M	2100 CT	ours 🌟	weights

Examples of fine-tuning our SuPreM on other downstream medical tasks are provided in this repository.

This work was supported by the Lustgarten Foundation for Pancreatic Cancer Research and the McGovern Foundation. The codebase is modified from NVIDIA MONAI. Paper content is covered by patents pending.