🏥 MedTri

MedTri: A Platform for Structured Medical Report Normalization to Enhance Vision–Language Pretraining

MedTri is a lightweight, locally deployable platform for structured normalization of radiology reports.
It transforms heterogeneous free-text reports into unified, anatomically grounded triplets, improving data consistency and fine-grained alignment for medical vision–language pretraining (VLP).

Paper: MedTri: A Platform for Structured Medical Report Normalization to Enhance Vision–Language Pretraining

---

📌 Overview

Medical vision–language pretraining (VLP) depends heavily on paired image–report data.
However, raw radiology reports commonly exhibit:

• Stylistic heterogeneity across institutions and radiologists
• Inconsistent length and verbosity
• Image-irrelevant content (e.g., clinical history, recommendations)
• Weak fine-grained image–text alignment

Existing normalization approaches typically:

• Focus on limited schema extraction (e.g., RadGraph-style NER pipelines), or
• Rely on cloud-based LLM rewriting, raising privacy and scalability concerns

---

🚀 What MedTri Provides

MedTri offers a structured, anatomy-grounded normalization framework that is:

• ✅ Lightweight
• ✅ Privacy-preserving (fully local deployment)
• ✅ Scalable for large-scale pretraining
• ✅ Designed for fine-grained image–text alignment

---

🧠 Core Concept: Structured Triplet Normalization

MedTri converts free-text reports into a unified schema:

[Anatomical Entity: Radiologic Description + Diagnosis Category]

---

🛠️ Usage

🔹 1. Report Normalization

Use the normalization script:

python perform_normalization.py

Input format: .xlsx radiology reports

Optional augmentation modules:

python explanation_augment.py   # MedTri-K
python counterfact_augment.py   # MedTri-C

---

🔹 2. Model Training

To train MedTri-based models:

python MedTri_train.py

A complete runnable example is provided in:

/example_train

---

📦 Model Release

Now you can freely download the pretrained checkpoints through the Google Drive link and update the path in perform_normalization.py to point to your local checkpoint directory.

---

🔬 Design Principles

MedTri is developed under three core principles:

1. Anatomy-grounded structuring instead of free-form rewriting
2. Local deployment instead of cloud dependency
3. Pretraining-oriented alignment instead of surface-level extraction

---

🙏 Acknowledgement

This project references and builds upon:

• RadGraph
• MedFILIP

We acknowledge their open-source contributions to the medical AI community.