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
Image Preprocessing Pipeline
Original images were provided in raw DICOM format. Each image was converted to PNG with the following preprocessing steps:
-
Outlier clipping
- The top and bottom 0.5% of DICOM pixel values were clipped
- Purpose: eliminate extremely dark or bright pixel outliers
-
Intensity normalization
- DICOM pixel values were linearly scaled to the 0–255 range
-
Resizing
- Images were resized so that the shorter side is 2048 pixels
- This was done to comply with Kaggle dataset size limits
-
View classification
- Each image was manually classified into:
- Frontal chest X-ray
- Lateral chest X-ray
