|
|
--- |
|
|
license: cc-by-nc-sa-3.0 |
|
|
--- |
|
|
|
|
|
This repo contains the checkpoints for OminiAbnorm-CT, a multi-modal generative model for grounded abnormality analysis on CT images from multiple planes and all human body regions. |
|
|
It supports three representative tasks: |
|
|
- Visual prompted report generation: Interpret an abnormality marked by a red bounding box, ellipse, contour, or cropped region. |
|
|
- Grounded report generation: Ground and interpret all abnormalities on the CT image. |
|
|
- Text-guided grounded report generation: Detect, ground and interpret a specific abnormality on the CT image. |
|
|
|
|
|
 |
|
|
|
|
|
It is built on **[OminiAbnorm-CT-14K](https://huggingface.co/datasets/zzh99/OminiAbnorm-CT-14K)**, the first large-scale dataset designed for abnormality grounding and description on multi-plane whole-body CT imaging. |
|
|
It contains 14.5K CT images with grounding annotation for 19K abnormal findings. |
|
|
Each abnormal finding is further linked to the detailed description in the report, and categorized according to a comprehensive hierarchical taxonomy. |
|
|
|
|
|
 |
|
|
|
|
|
Check our [paper](https://www.arxiv.org/abs/2506.03238) and [github repo](https://github.com/zhaoziheng/OminiAbnorm-CT) for usage and more details. |
