Detect At-Risk Hips from Pelvic X-rays
RiskRay is a research tool for screening hip fracture risk from DICOM pelvic X-rays using a mixed-data convolutional neural network. Upload a scan, crop the left and right hips, and get a model-based risk score in seconds.
How It Works
RiskRay combines DICOM image processing with a mixed CNN/MLP architecture to score hip fracture risk from pelvic radiographs.
Upload DICOM
Drag and drop a DICOM or image file
Supports standard DICOM files and common image formats. The browser-side DICOM viewer renders a preview so you can verify the scan before proceeding.
Crop the Hips
Annotate left and right hip regions
Draw bounding boxes around the left and right femoral heads. These crops are extracted server-side and passed to the model as individual inputs.
Get Risk Scores
Mixed CNN inference per hip
The mixed-data CNN scores each cropped hip independently. Results are stored with the original scan metadata for later review and comparison.
Capabilities
Built for research and iterative model improvement.
Inference
Upload → crop → predict → store
Run inference on new pelvic X-rays. Each upload is saved to an S3-backed store with its crop coordinates and per-hip prediction scores, building a queryable history of screening results.
Training Pipeline
Async model training and testing jobs
Submit training jobs with configurable epochs, learning rate, batch size, and validation split. Jobs run asynchronously via FastAPI BackgroundTasks and report status back to the UI.
Upload Library
Searchable history of all scans
Browse past uploads by patient name, date, or purpose. View detail pages with full images, crop overlays, prediction scores, and clinical notes.
Mixed-Data CNN
Image + metadata architecture
The model uses a dual-branch architecture: a CNN processes the cropped hip image while an MLP branch ingests DICOM metadata. The concatenated features feed into dense layers for the final risk score.
Start Screening
Upload a pelvic X-ray, annotate the hips, and see the model's risk assessment.