Skip to content

MODEL_CARD.md

Latest commit

 

History

History
83 lines (56 loc) · 3.54 KB

MODEL_CARD.md

File metadata and controls

83 lines (56 loc) · 3.54 KB
Raw

PathoSynVLM Model Card

Paper | Code | Model weights | Training data | Project page

Model

PathoSynVLM is a token-efficient vision-language model for case-level pathology synoptic report generation.

Architecture:

  • Frozen WSI patch encoder: CONCHv1.5
  • Vision-language bridge: two-layer MLP aligner
  • Language decoder: Qwen2.5-3B-Instruct
  • Case structure: one or more WSI embedding files with optional WSI marker tokens

Intended Use

The model is intended for research on pathology report generation from precomputed WSI patch embeddings. Given one or more slide embedding files for a case, it generates:

Diagnosis: ...
Certainty: ...
Conclusion: ...

This model is not a clinical diagnostic device and should not be used for patient care without appropriate validation, regulatory review, and expert oversight.

Training Data

The paper-relevant pipeline uses:

The official HISTAI metadata repository links to its organ-specific WSI repositories. See the HISTAI source documentation for dataset structure, access instructions, licensing, and citation. Users are responsible for following each dataset's access terms and redistribution rules.

Released Weights

The official release is hosted at AtlasAnalyticsLab/PathoSynVLM. Download it into:

$PATHOSYNVLM_WEIGHTS_ROOT/pathosynvlm-stage2-main/

source configs/paths.example.env
hf download AtlasAnalyticsLab/PathoSynVLM \
  --local-dir "$PATHOSYNVLM_WEIGHTS_ROOT/pathosynvlm-stage2-main"

See docs/weights.md for the required package layout. The Stage 2 paper run used unfreeze_llm_base=true, so the release package includes the merged/full LLM weights, not only a LoRA adapter.

Reported Metrics

The main reported Stage 2 HISTAI result is:

ROUGE-L METEOR BLEU-4 BERTScore F1 Diagnosis Exact Diagnosis Relaxed Certainty
0.2495 0.1988 0.0525 0.3018 0.1667 0.3333 0.9000

See configs/reported_results.json for the machine-readable values.

Limitations

  • Inputs are precomputed patch embeddings, not raw WSIs.
  • Result matching depends on using the same CONCHv1.5 feature format and metadata filtering.
  • The model may generate incorrect or incomplete pathology statements.
  • Dataset distributions and reporting styles may not generalize across institutions.

Citation

Yang, Z.; Cheng, J.; Trinh, V. Q.-H. and Hosseini, M. S. (2026). Simple Token-Efficient Vision-Language Model for Case-Level Pathology Synoptic Report Generation. In Proceedings of the 7th International Conference on Deep Learning Theory and Applications, ISSN 2184-9277, pages 514–537.

@inproceedings{yang2026simpletokenvlm,
  title     = {Simple Token-Efficient Vision-Language Model for Case-Level Pathology Synoptic Report Generation},
  author    = {Yang, Zhiyuan and Cheng, Jiahao and Trinh, Vincent Quoc-Huy and Hosseini, Mahdi S.},
  booktitle = {Proceedings of the 7th International Conference on Deep Learning Theory and Applications},
  pages     = {514--537},
  year      = {2026},
  issn      = {2184-9277}
}