| license: apache-2.0 | |
| # PET2Rep | |
| [**📖 arXiv**](https://arxiv.org/abs/2508.04062) | [**🎈 GitHub**](https://github.com/YichiZhang98/PET2Rep) | [**🤗 HuggingFace**](https://huggingface.co/datasets/r1pples/PET2Rep) | |
| ## Overview | |
| **PET2Rep** is the first benchmark dataset designed for the evaluation of VLMs in the context of Positron Emission Tomography (PET) radiology report generation. | |
| This dataset comprises 565 whole-body PET/CT image-report pairs. | |
| It offers extensive coverage of dozens of organs with corresponding metabolic information, which is crucial for reflecting the real-world clinical comprehensiveness of PET imaging in oncologic diagnostics. | |
| The primary motivation behind the creation of PET2Rep is to facilitate the development and standardized evaluation of automated systems for generating clinically accurate and relevant radiology reports from PET/CT scans, thereby addressing a significant gap in current medical imaging datasets. | |
| ## Data Fields | |
| The dataset contains the following fields: | |
| * `image_data`: Whole-body PET/CT scan images in JPEG format. | |
| The structure of JPEG file: **subject_\<subject id\>_\<slice\>_coronal.jpg** | |
| * `general_prompt`: A prompt file consists of key-value pairs for the radiology report and other supplementary information in JSON Lines format. |