| This repo contains the checkpoints for SAT. | |
| We offer SAT-Pro, SAT-Nano (both trained on 72 datasets) and another 5 different variants of SAT-Nano (all trained on 49 datasets): | |
| - SAT-Pro: ./Pro | |
| - SAT-Nano: ./Nano | |
| - UNET-Ours: ./Others/UNET-Ours | |
| - UNET-CPT: ./Others/UNET-CPT | |
| - UNET-BB: ./Others/UNET-BaseBERT | |
| - UMamba-CPT: ./Others/UMamba-CPT | |
| - SwinUNETR-CPT: ./Others/SwinUNETR-CPT | |
| Check our [paper](https://github.com/zhaoziheng/SAT/tree/main) for more details, and [github repo](https://github.com/zhaoziheng/SAT/tree/main?tab=readme-ov-file) for usage instruction. | |
| ⚠️ Each model should be used with paired checkpoint and text encoder checkpoint. | |
| In addition, we provide multiple pretrained encoders at ./Pretrain. Enhanced with multi-modal human anatomy knowledge, they significantly boost the segmentation performance and are potentially beneficial for other tasks: | |
| - A version pretrained only with the textual knowledge (`textual_only.pth`). | |
| - A version further pretrained with [SAT-DS](https://github.com/zhaoziheng/SAT-DS/tree/main) (`multimodal_sat_ds.pth`). It can be used to reproduce results in our [paper](https://arxiv.org/abs/2312.17183). | |
| - A version further pretrained with 10% training data from [CVPR 2025: FOUNDATION MODELS FOR TEXT-GUIDED 3D BIOMEDICAL IMAGE SEGMENTATION](https://www.codabench.org/competitions/5651/) (`multimodal_cvpr25.pth`). It's explicitly optimized for the challenge. | |