← INDEX/WORK/VERIDOSE
№ 03 — RADIATION-ONCOLOGY AI
Anyone can ship an auto-contour model. The question is: who signs their name to it?
THESIS
In radiation therapy, clinicians hand-trace every organ on every CT slice — the contour is the treatment plan. It’s slow, and it fails exactly where it matters most: the prostate/bladder/rectum region has almost no visible density edge for classical tools to find.
Veridose’s answer is two-sided. Technically: AI contouring built on learned 3D shape priors rather than pixel intensity. Commercially: accountability as the moat — every model release reviewed and cryptographically signed by a named, board-certified medical physicist (my co-founder, Maurice Tajiran, MS, DABR), with public per-anatomy benchmarks and a tamper-evident audit trail.
THE BUILD
FastAPI service running TotalSegmentator (nnU-Net v2) on staged CT cases with GPU inference on an RTX 5070 Ti. Per-anatomy NIfTI masks stream into a multi-label volume one file at a time — holding RAM near 80 MB instead of ~3 GB.
On a real pelvis CT (public TCIA data — no PHI ever committed), the pipeline measured that ~80% of soft-tissue voxels have sub-threshold gradient: there is literally no edge to segment. That number is the thesis — shape priors or nothing.
A NiiVue-based multiplanar viewer overlaying ten segmented structures on the CT with per-structure colormaps, opacity controls, slice scrubbing, and axial / coronal / sagittal / 3D modes.
A hash-chained, tamper-evident ledger for linac calibration attestations (TG-51 / TG-142). Flip a single byte anywhere in history and the chain visibly breaks; validation runs on every read. The signature-as-moat thesis, embodied in code.
A from-scratch GLSL fragment shader — FBM tissue grain, a sweeping scan line, converging dose beams, pulsing isodose rings — with a 30 fps cap, DPR clamping, IntersectionObserver pausing, context-loss recovery, and full reduced-motion fallback.
Next.js 16 / React 19 / Tailwind v4 with bespoke effects (no component library), NASA-blue trust palette, monospace verification motifs, and honest labeling: every illustrative number marked as such.
SCREENS
SIGNALS
STACK
WHAT IT PROVES
That I can walk into a regulated, unforgiving domain — medical physics — and ship a working end-to-end system: real GPU inference, real 3D visualization, real cryptographic audit machinery, and marketing that respects the line between demo and clinic.