| --- |
| license: cc-by-nc-4.0 |
| language: |
| - en |
| pipeline_tag: image-to-text |
| tags: |
| - medical |
| - pathology |
| - vision-language |
| - contrastive-learning |
| - fine-grained |
| - multimodal |
| library_name: transformers |
| --- |
| # PathFLIP |
|
|
| Model weights for the paper *PathFLIP: Fine-Grained Language-Image Pretraining for Versatile Pathology Image Understanding*. |
|
|
| ## Overview |
|
|
| PathFLIP is a pathology vision-language model that aligns fine-grained morphological sub-captions with their corresponding regions in Whole Slide Images. Unlike prior pathology VLMs that pair an entire slide with a single report-level anchor, PathFLIP introduces region-statement correspondence through a region Q-Former and a region-level contrastive objective with caption-swapped negatives, learning region-level alignment without any manual spatial annotation. This fine-grained supervision enables strong slide-level classification and retrieval performance, and gives rise to an emergent visual grounding capability. |
|
|
| ## Model Details |
|
|
| - **Base model**: *Qwen3-0.6B* |
| - **Training data**: [FGC-4K Dataset](https://huggingface.co/datasets/jshhhh/PathFLIP/) |
| - **Task**: classification, image-text retrieval, visual grounding, vqa |
| - **Languages**: English |
|
|
| ## License |
|
|
| This model is released under CC BY-NC 4.0 — free for academic and research use, **not for commercial use or clinical deployment**. |
|
|
