RT-Super trains tumor segmentation from radiology reports, lifting Dice from 16 to 30 without masks
maier_ak · x · 2026-10-13
A preprint from Johns Hopkins, Harvard Medical School, UCSF and others introduces RT-Super (code public), replacing costly hand-drawn masks with routinely available data: radiology reports, longitudinal scans, and multi-phase CT.
Teacher-student self-distillation
- The student, a MedFormer segmentation net, sees one CT and no report at deployment
- The hybrid conv+transformer teacher also reads the report, other scans and contrast phases
- A report-aware transformer generates conv kernels to refine student features; where no mask exists, the student mimics the teacher
Results and limits
- Detection is a tie with predecessor R-Super: external AUC 84 for both
- The real gain is mask-free outlining: Dice 16 → 30
- Far from the ceiling: two radiologists agree at Dice 50–66
Andreas Maier notes the irony: the report, once seen as a privacy leak, is now the supervision.
More from Research
- AgentSWE benchmark: no coding agent comes close to delivering the agents it's asked to build — Jiahao Wang · 2026-10-13
- Schema agent hits ~99% RHAE on ARC-AGI-3 public set, full paper and code released — wzenus · 2026-10-13
- World model research: more data won't suffice—squeeze more from it; even intuitive physics is still out of reach — mangahomanga · 2026-10-13
- Indie dev builds own serving library to train SVG model, finds evolution loses to SGD — cephaloform · 2026-10-13
- Harvard & NTT Research: cooperating AI agent populations hit an Allee-effect takeoff threshold — Hidenori8Tanaka · 2026-10-13
- Princeton's QUEEN: a 4B model that plays grandmaster-level chess and explains its moves in plain language — jiqizhixin · 2026-10-13