Upload a chest X-ray and receive AI-powered detection, multi-class classification, and explainable Grad-CAM heatmaps — trained on 5,000+ VinDr-CXR images.
Trained on the VinDr-CXR dataset to identify and localize a wide range of pulmonary conditions.
Simultaneously detects multiple co-occurring pathologies in a single X-ray. The model outputs calibrated probability scores per class, enabling nuanced clinical triage rather than binary detection.
YOLOv8m draws precise localization boxes over abnormal regions with class labels and confidence scores for each finding.
Gradient-weighted class activation maps overlay the X-ray to visualize exactly which image regions influenced the AI decision.
Optimized model pipeline delivers results in under 0.4 seconds using TorchScript-compiled weights and ONNX export.
All models evaluated on AUC-ROC, F1-score, and mAP metrics. Comprehensive reports generated per training run for reproducibility.
Production-grade ML stack from data ingestion to Streamlit deployment.
An 8-week structured pipeline from raw data to deployment-ready system.
Download VinDr-CXR, develop DICOM-to-PNG conversion scripts, parse and convert CSV annotations to YOLO/COCO format, and validate the full preprocessing pipeline.
Hyperparameter tuning with AdamW vs SGD, advanced Albumentations augmentation, transfer learning, Grad-CAM implementation, and full validation set evaluation.
Integrate EfficientNet/ResNet for classification and YOLOv8m for detection, implement data loaders and augmentation pipelines, train baseline models, and report initial AUC and mAP scores.
Final evaluation on test set, complete documentation and README, Streamlit web interface with best.pt weights loaded, ONNX export, and clinical readiness assessment.
Upload a chest X-ray (JPG/PNG) and get instant AI-powered detection results from your trained YOLOv8 model.
Make sure app.py is running in your terminal first.