yohanaraphael19/kansa-vision

AI-assisted breast cancer screening for low-resource settings # KansaVision > AI-assisted breast cancer screening for low-resource clinical settings. KansaVision is a private web application for hospital radiologists. A doctor uploads a breast ultrasound scan; the system runs a local AI pipeline and produces a structured report with a **BI-RADS category**, confidence score, and segmentation overlay. The doctor reviews, optionally overrides, and signs off. Every action is audit-logged. > **Kansa** is the Swahili word for *cancer*. The name reflects the project's > East African roots and its focus on hospitals like Katavi Regional Referral > Hospital, Tanzania — where radiologists are scarce but diagnostic need is real. > **Disclaimer:** Clinical Decision Support — not a diagnosis. Requires > radiologist review and sign-off. --- ## Pipeline Upload scan → Denoising → Enhancement → Segmentation → BI-RADS Classification → Clinical Report --- ## Stack | Layer | Choice | |---|---| | Frontend | React 18, Vite, React Router v7, CSS Modules | | Backend | FastAPI, SQLAlchemy, Alembic, JWT | | Database | SQLite (dev) / PostgreSQL (prod) | | ML | PyTorch, OpenCV, SAM 2, MedGemma 4B | ### Models | Task | Model | |---|---| | Denoising + enhancement | Bilateral filter + CLAHE (OpenCV) | | Lesion segmentation | MedSAM 2 | | BI-RADS classification + report | MedGemma 4B via HuggingFace Spaces API | Segmentation weights are cached locally. BI-RADS classification is handled by a MedGemma 4B instance hosted on HuggingFace Spaces — requires an active Space and API access. --- ## Access Model | Role | Capabilities | |---|---| | **Admin** | Register/deactivate doctors, view all scans, manage settings | | **Doctor** | Upload scans, run analysis, review findings, sign reports | - First registered user auto-promotes to admin. Self-signup is disabled thereafter. - Every scan stores: `doctor_id`, `timestamp`, `model_version`, `bi_rads_output`, `doctor_override` — full audi …