RadiAid
Technological Framework & Systems
Proof-of-Concept
Our binary classifier uses an attention condenser network, a type of convolutional neural network that's optimized for pattern recognition in medical images.
Current Performance:
- 93.75% accuracy
- 341,000 parameters
- Under 2MB storage
- Runs in under 1 second on mid-tier smartphones
Training Process
The model was trained on 442 bone X-rays and tested on 96 images. It detects patterns like:
- Irregular bone texture
- Abnormal mineralization
- Periosteal reactions characteristic of osteosarcoma
Adding patient metadata (age, gender) increased accuracy from 83.33% to 93.75%.
Standardized Image Capture
The quality box attachment ensures consistent, high-quality X-ray photos:
- High CRI LED panels provide uniform back-illumination
- Anti-reflective surface prevents glare
- Fixed-distance mount ensures consistent capture angle
- Transforms any smartphone into a precision X-ray scanner
Building Trust
We use Grad-CAM heatmaps to show which regions of the X-ray influenced the AI's decision. This transparency helps clinicians understand and trust the system's reasoning
We also employ Tversky Loss function to penalize false negatives more heavily than false positives making sure that we rarely miss malignant cases.