lalainaandrianina/medical-report-ai

Generative AI for Automated Medical Report Generation — Deep Learning applied to Chest X-Ray Medical Imaging — AIMS Rwanda 2026 # Generative AI for Automated Medical Report Generation A deep learning system that automatically generates medical reports from chest X-ray images, specialized for the African healthcare context. ## Project Overview This project addresses the critical shortage of radiologists in Africa by automating medical report generation using generative AI. The system is trained on 11,731 paired clinical reports and achieves a ROUGE-1 score of 0.53. ## Features - Text-based report generation using T5-small - TB-specialized African model (Val Loss: 0.07) - Direct X-ray image to report pipeline (ResNet50 + T5) - Multilingual output: English and French - Web interface built with Streamlit ## Datasets - Indiana University Chest X-ray Dataset (Kaggle) - TBX11K Tuberculosis Dataset (Kaggle) - Combined: 11,731 training samples ## Results | Metric | Before | After | |--------|--------|-------| | ROUGE-1 | 0.418 | 0.529 | | ROUGE-2 | 0.316 | 0.427 | | ROUGE-L | 0.416 | 0.524 | ## Mathematical Framework - Conditional sequence modeling: p_theta(y | x) - Empirical Risk Minimization - Cross-entropy loss - AdamW optimization (lr = 5e-5) ## Installation ```bash pip install transformers torch torchvision streamlit rouge-score sentencepiece Pillow pandas scikit-learn ``` ## Run the Web Interface ```bash streamlit run medical_report_app.py ``` ## Author Zo Lalaina Andrianina Andrianantenaina MSc Mathematical Sciences - Epidemiology AIMS Rwanda | Madagascar lalaina.andrianina@aims.ac.rw ## Acknowledgements This project was developed independently alongside my Master's thesis at AIMS Rwanda, using publicly available datasets from Kaggle.