AI-Powered Radiology Report Generator
Clinical Decision Support API for chest X-rays using DenseNet-169 + Grad-CAM with GPT-5 report generation. Provides predictions, heatmaps, and clinician/patient-friendly summaries via REST endpoints.
Core trick
DenseNet-169 classifies 14 chest X-ray conditions with Grad-CAM heatmaps, and GPT-5 turns the predictions into clinician or patient reports.
Description
This project implements a Clinical Decision Support (CDS) API for automated chest X-ray interpretation. Built on a pretrained DenseNet-169 model fine-tuned with the ChestX-ray14 dataset, it detects 14 common thoracic conditions plus “No Finding.” Grad-CAM is integrated to generate heatmaps, allowing developers and clinicians to visualise the regions that most influenced predictions. To bridge the gap between raw model output and clinical usability, GPT-5 is used to generate human-readable reports, configurable for either clinicians (technical, diagnostic detail) or patients (simplified explanations). The API is built with FastAPI and exposes endpoints for prediction, Grad-CAM overlay retrieval, and report generation. Key benefits include explainable predictions, modular design for integration into research or healthcare applications, and support for multimodal workflows combining imaging AI with natural language generation.
README excerpt
# Radiology CDS (DenseNet-169 + Grad-CAM + GPT-5) A **Clinical Decision Support (CDS)** system for chest X-ray analysis using **deep learning (DenseNet-169), explainability (Grad-CAM), and GPT-5** to generate clinician-friendly and patient-friendly diagnostic reports. <video src="https://github.com/mazshuky/Radiology-Report-Generator/blob/main/demo.mp4" width="600" controls> </video>  ## Features - Multi-label classification of 14 thoracic diseases + _No Finding_ - Grad-CAM overlays for visual explanation of model predictions - GPT-5 powered reporting (patient-friendly or clinician-friendly) - REST API (FastAPI) for integration into clinical workflows - React UI for uploading images, viewing overlays, and reading reports - Built-in disclaimer for safe research use ## Tech Stack - **Backend**: Python, FastAPI, PyTorch, OpenAI GPT-5 - **Frontend**: React, Tailwind CSS - **Model**: DenseNet-169 pre-trained on [NIH Chest X-rays](https://www.kaggle.com/datasets/nih-chest-xrays/data) dataset ## Project Structure ``` ├── requirements.txt ├── densenet169_chestxray14.pth ├── preprocess_data.py ├── train.py ├── evaluate.py ├── radiology_cds_api.py ├── assets/ │ ├── class_distribution.png │ ├── confusion_matrices.png │ └── roc_curves.png ├── dataset/ │