Cell Bound

Try both live versions

• Original green-and-cream Fruit-stand demo: Open the original version
• Current pixel-laboratory demo: Open the current version

Inspiration

Bone-marrow cellularity estimation is visually demanding and specialized digital-pathology tools are not always available. Cell Bound explores a transparent, local-first teaching workflow that helps trainees reason about occupied area without pretending to diagnose disease.

What it does

Cell Bound analyzes one representative H&E field locally in the browser. A deterministic engine separates heuristic image compartments, groups recurring visual appearances, reports up to five groups plus Other against one fixed ROI denominator, and lets users reversibly subtract groups without re-clustering or renormalizing. It includes a project-authored pixel-fruit primer and synthetic marrow-like demo, so judges can try the complete workflow without patient data.

How we built it

The app uses React, TypeScript, vinext/Vite, Canvas APIs, and Cloudflare Workers-compatible hosting. Codex with GPT-5.6 translated the concept into the interface, deterministic image-analysis pipeline, API boundary, adversarial tests, documentation, and deployment workflow. GPT-5.6 also helped design and test an optional constrained summary path that receives only validated derived metrics and quality flags—never image pixels, filenames, or metadata. The measured result always comes from the deterministic engine; a local fallback keeps the analyzer usable without an API key.

Challenges

The hardest parts were stable cluster ordering, exact fixed-denominator accounting, reversible subtraction, honest clinical boundaries, browser performance, and preventing image-like data from reaching the optional model route.

Accomplishments

We built a free no-login workflow, deterministic masks and arithmetic, a visible audit equation, strict privacy/model boundaries, responsive controls, and 18 passing automated tests covering pathological inputs, determinism, API validation, and fallback behavior.

What we learned

Transparent intermediate states matter more than a single impressive score. Stable ordering and integer area bookkeeping are essential for reproducibility, and a narrow AI role can be more useful and safer than an opaque end-to-end prediction.

What's next

Future work starts with governed evaluation on approved de-identified fields, separate tuning and test sets, broader device benchmarking, manual correction tools, and multi-field comparison. Cell Bound is a research and education proof of concept, not a medical device or replacement for professional judgment.

Image credits

Slide-viewer imagery shown in the demo video is © University of Leeds and comes from the University of Leeds Virtual Pathology Project. It is included for educational demonstration only; no affiliation or endorsement is implied. These third-party screenshots are not bundled with Cell Bound or used by either live demo, which use project-authored synthetic imagery. See the Leeds Virtual Pathology terms.

This image was originally published in ASH Image Bank. Reva Channah Goldberg; Girish Venkataraman. Blasts on the Bone Marrow Biopsy. ASH Image Bank. 2019; image 62449. © the American Society of Hematology. See the ASH Image Bank permissions and terms. No author or ASH endorsement is implied.

Built With

• canvas
• cloudflare
• openai