Medal S V1.0

--- pipeline_tag: image-segmentation license: apache-2.0 --- # Medal S: Spatio-Textual Prompt Model for Medical Segmentation [](https://arxiv.org/abs/2511.13001) [](https://openreview.net/forum?id=9vCx66pnLn#discussion) [](https://huggingface.co/spc819/Medal-S-V1.0) [](https://github.com/yinghemedical/Medal-S) This repository provides guidance for training and inference of Medal S within the [CVPR 2025: Foundation Models for Text-Guided 3D biomedical image segmentation](https://www.codabench.org/competitions/5651/) Docker link for the 2025/05/30 testing submission: [Medal S](https://drive.google.com/file/d/1HRJqYUXajptGsKaXEhn-s3rGcnKIwGs7/view) ## Requirements **Python Version:** 3.10.16 ### Installation 1. **Create conda environment and install dependencies:** ```bash #!/bin/bash # Create environment conda create -n medals_local_test python=3.10 -y conda activate medals_local_test # Install packages pip install torch==2.2.0 torchvision==0.17.0 torchaudio==2.2.0 --index-url https://download.pytorch.org/whl/cu121 pip install transformers==4.51.3 monai==1.4.0 nibabel==5.3.2 tensorboard einops positional_encodings scipy pandas scikit-learn scikit-image batchgenerators acvl_utils # Install nnU-Net wget https://github.com/MIC-DKFZ/nnUNet/archive/refs/tags/v2.4.1.tar.gz tar -xvf v2.4.1.tar.gz pip install -e nnUNet-2.4.1 # Install dynamic network architectures cd model && pip install -e dynamic-network-architectures-main && cd .. ``` ## Training Guidance First, download the dataset from [Hugging Face: junma/CVPR-BiomedSegFM](https://huggingface.co/datasets/junma/CVPR-BiomedSegFM). * **Data Preparation**: Preprocess and organize all training data into a `train_all.jsonl` file using the provided script: `data/challenge_data/get_train_jsonl.py`. * **Knowledge Enhancement**: You can either use the pre-trained text encoder from SAT ([https://github.com/zhaoziheng/SAT/tree/cvpr2025challenge](https://github.com/zhaoziheng/SAT/tree/cvpr2025challenge)) available on [Hugging Face](https://huggingface.co/zzh99/SAT/tree/main/Pretrain), or pre-train it yourself following the...