Bittensor subnet for SOTA radiology pre-screening AI
The SotaRad subnet is a radiology pre‑screening / triage subnet on Bittensor designed to push forward practical, clinically useful chest imaging models. It focuses on detecting a small but critical set of conditions – currently Tuberculosis, Pneumonia, Bronchitis, and Silicosis – and rewards miners for models that generalize well to new, time‑shifted real‑world data, not just a fixed benchmark.
Miners don’t stream per‑request predictions; instead they publish full models (Hugging Face repo + revision) and either host them on Chutes or expose them via a local OpenAI‑compatible SGLang server. Validators read on‑chain commitments, spin up the corresponding model endpoints, and then run a standardized vision‑language prompt to classify chest radiology studies. Each sample is scored with a simple but robust flag‑only scheme: if the model outputs any structured “finding” for a target condition, that counts as a positive; otherwise it is negative. This yields TP/FP/FN/TN counts per miner and time period, from which the subnet computes precision, recall, and an Fβ score (with β>1β>1 to overweight recall and penalize false negatives).
A core design feature is temporal integrity: every model is only evaluated on studies acquired after its on‑chain commit time plus a configurable evaluation delay (e.g. ~1 day in production). This ensures miners cannot pre‑train directly on the exact evaluation samples they will later be judged on, and instead are rewarded for true generalization to future data. Evaluation runs in discrete time buckets (evaluation periods); validators store per‑UID scores in SQLite, and then aggregate across recent periods to infer stable performance rather than reacting to single noisy batches.
Incentives are implemented via a tiered, winner‑takes‑most emission mechanism. Several tiers (A–E) each look back over a different number of recent periods, select either the top‑1 miner or a top percentage, and assign a fixed share of the emission budget. Within each tier, miners are ranked primarily by mean Fββ, then recall, then precision, and finally by model size (fewer parameters preferred) and earlier commit as tie‑breakers. The top tier concentrates most of the emission on the best performer, while lower tiers provide smaller rewards to other consistently strong miners. The resulting per‑UID emissions are combined and normalised into weights that the validator pushes on‑chain.
Overall, this subnet is designed to be:
Clinically grounded: focused on a small, high‑impact target condition set with clear utility in triage and pre‑screening, not diagnosis replacement.
Generalization‑driven: rewards are based on future‑period performance under time delay, not static leaderboard tuning.
Model‑centric: miners compete with whole model checkpoints served behind standard OpenAI‑style APIs, making deployment and audit straightforward.
Strongly competitive but stable: the tiered, time‑windowed reward design behaves like a soft winner‑takes‑all, concentrating incentives on frontier models while still tracking performance over multiple recent periods to reduce noise.
In short, the SotaRad subnet turns continuous real‑world radiology data into a live, time‑shifted benchmark and uses Bittensor’s incentive layer to drive an open competition for the best chest imaging pre‑screening models.
Data acquisition agreement procured, testing successful, looking at TEE at future solution for more advanced imaging cases
Investment agreements in place, conditional upon subnet launch