MedMetric is a subnet on Bittensor designed to produce a single commodity: production-grade, edge-deployable AI models for medical imaging analysis — chest X-ray, brain MRI, and bone CT.
Miners compete globally to build the best imaging segmentation models, scored on clinical accuracy, edge inference speed, and out-of-distribution robustness. Validators maintain curated clinical benchmark datasets and objectively evaluate miner submissions through Yuma Consensus. Hospital customers purchase top-ranked models and deploy them locally on edge hardware (NVIDIA Jetson AGX Orin, Intel NUC, DICOM workstations) — with zero data egress and full regulatory compliance.
The result is a decentralized R&D engine for clinical AI: global competition between miners continuously produces improving models, and real hospital demand steers development toward what medicine actually needs.
Why it matters:
Score = 0.50 x ClinicalAccuracy + 0.30 x EdgeSpeed + 0.20 x OOD_Robustness
| Dimension | Weight | What It Measures |
|---|---|---|
| Clinical Accuracy | 50% | Segmentation quality vs. ground-truth annotations (Dice + BCE) |
| Edge Speed | 30% | Wall-clock inference time on Jetson Orin NX 8GB reference hardware |
| OOD Robustness | 20% | Prediction consistency under augmented/noisy inputs |