thaiphonghuan/BLIP-finetuned-chest-xray-v1

Fine-tuned BLIP on Chest X-rays (Indiana University)

This repository contains a fine-tuned BLIP (Bootstrapped Language-Image Pretraining) model trained on the Chest X-rays (Indiana University) dataset.
The model is adapted for vision–language tasks in the medical imaging domain, particularly chest X-ray understanding.

---

🧠 Model Description

• Base model: BLIP (Bootstrapped Language-Image Pretraining)
• Fine-tuning domain: Medical imaging
• Modality: Vision–Language (Image + Text)
• Target data: Chest X-ray images (frontal & lateral views)

The goal of fine-tuning is to adapt BLIP to better capture radiological visual patterns and associated semantic information from chest X-ray images.

---

📊 Dataset Information

The model is fine-tuned using the Chest X-rays (Indiana University) dataset.

Dataset Source

• Kaggle:
  https://www.kaggle.com/datasets/raddar/chest-xrays-indiana-university

Image Preprocessing Pipeline

Original images were provided in raw DICOM format. Each image was converted to PNG with the following preprocessing steps:

1. Outlier clipping

   • The top and bottom 0.5% of DICOM pixel values were clipped
   • Purpose: eliminate extremely dark or bright pixel outliers

2. Intensity normalization

   • DICOM pixel values were linearly scaled to the 0–255 range

3. Resizing

   • Images were resized so that the shorter side is 2048 pixels
   • This was done to comply with Kaggle dataset size limits

4. View classification

   • Each image was manually classified into:
     • Frontal chest X-ray
     • Lateral chest X-ray