MedTri Checkpoints and Training Data for Structured Medical Report Normalization in Vision–Language Pretraining

MedTri is a lightweight, locally deployable framework for structured normalization of radiology reports, designed to improve medical vision–language pretraining by converting heterogeneous free-text reports into anatomy-grounded triplets. The normalized representation follows the schema:

[Anatomical Entity: Radiologic Description + Diagnosis Category]

This Zenodo record provides the pretrained MedTri checkpoint and associated training data used for structured medical report normalization. The released files are intended to support reproducible deployment, evaluation, and further development of MedTri-based report normalization pipelines.

MedTri addresses several common limitations of raw radiology reports used in vision–language pretraining, including stylistic variability, inconsistent report length, image-irrelevant content, and weak fine-grained image–text alignment. Instead of relying on cloud-based large language model rewriting, MedTri is designed for local deployment, making it suitable for scalable and privacy-preserving preprocessing of medical image–report datasets. The GitHub repository describes MedTri as a lightweight local platform that transforms free-text radiology reports into unified, anatomically grounded triplets for improved data consistency and fine-grained alignment.

The archive includes:

checkpoint_MedTri.zip: pretrained MedTri model checkpoint for local report normalization.

train.json: training data used for MedTri model development and reproduction.

This release is associated with the paper “MedTri: A Platform for Structured Medical Report Normalization to Enhance Vision–Language Pretraining.” The paper presents MedTri as a deployable normalization framework that converts free-text reports into unified anatomy-grounded triplets, preserving clinically relevant morphological and spatial information while reducing stylistic noise and image-irrelevant text. Across X-ray and CT datasets, the study reports that structured anatomy-grounded normalization improves medical vision–language pretraining compared with raw reports and existing normalization baselines.

Suggested citation

Chu,Y.,Ma,X.,Jin,X.,Luo,G.,&Gao,X.MedTri: A Platform for Structured Medical Report Normalization to Enhance Vision–Language Pretraining. arXiv:2602.22143,2026.

Related code

GitHub repository: Arturia-Pendragon-Iris/MedTri