This is the repository of BraTS-ReportX, a paired resource of 257 clinical reports aligned to BraTS subjects, structured into a rich set of qualitative and quantitative attributes. This repository contains code for:
- Automatic generation of quantitative report attributes from BraTS data, including anatomical localization and geometric measurements;
- Report encoding with biomedical language models;
- Evaluating the semantic coverage and overall quality of the dataset, supporting analyses of how well BraTS-ReportX captures clinically relevant report information compared with existing resources; and (4) training and testing of the proposed vision-text alignment framework for 3D tumor segmentation. The codebase is designed to support reproducibility and further research on integrating structured clinical semantics into medical image segmentation.
Overview of the annotation protocol. Clinician reports and automatically generated reports are produced independently and then concatenated.
Our segmentation pipeline overview. Encoder features from a 3D U-Net are projected into flat visual embeddings, while clinical reports are mapped to text embeddings. A contrastive vision-text module aligns both modalities during training, while inference relies only on the image backbone.
The project structure is available here.
- Clone the repository:
git clone https://github.com/AImageLab-zip/Report-Guided-Segmentation
cd Report-Guided-Segmentation- Create a virtual environment and install dependencies:
uv sync --no-cache- Activate the environment
source .venv/bin/activateThis is a fork of a universal framework from https://github.com/kev98/Medical-Image-Segmentation. An overview on using and extending it with your own implementation can be found here