Isaackjoshua/mwana-ai-ios

AI-assisted breast cancer screening iOS app (Flutter + on-device ONNX + Gemma 4) # Mwana-AI AI-assisted breast cancer screening iOS app. Analyses breast ultrasound images entirely on-device — no patient data ever leaves the phone. --- ## Screenshots Splash Screen Model Setup Original Ultrasound Bounding Box Overlay Grad-CAM Heatmap > Heatmap screenshots are rendered from a synthetic test ultrasound via `dart run tool/test_heatmap.dart`. Full-workflow screenshots (analysis, cumulative report) require the on-device ONNX and Gemma models. --- ## What It Does ### Single-Image Workflow 1. Pick a breast ultrasound image — Camera, Photo Library, Files app, or **Butterfly iQ probe** 2. On-device ONNX inference classifies the image as **Benign / Malignant / Normal** and segments the lesion boundary 3. An ACR BI-RADS category (1 / 2 / 3 / 4A / 4B / 4C–5) is assigned 4. The segmentation mask is rendered as a colour overlay with a bounding box 5. Tap the image to open the **full-screen viewer** — toggle between Original, Bounding Box, and Grad-CAM Heatmap with pinch-to-zoom 6. On-device Gemma 4 generates a structured clinical report (falls back to a deterministic template if the model is not loaded) 7. Export the report as a PDF via the iOS share sheet ### Multi-Section Probe Workflow When using the Butterfly iQ probe, the examiner captures **multiple breast sections** in a single session: 1. Connect the Butterfly iQ probe — the app authenticates automatically using the embedded client key 2. Capture Section 1, Section 2, … Section N (thumbnail strip shows all captured frames) 3. Tap **Analyse N** — the app runs sequential ONNX inference on every section 4. A **Cumulative Report** screen shows: - Per-section BI-RADS classifications with sortable confidence bars - Bounding box and heatmap thumbnails (tappable) for each section - An overall worst-case BI-RADS banner; **"Most Concerning"** section badge only appears when sections genuinely differ - AI-generated narrative from Gemma 4 covering all sections - Editable report fields (Clinical Indi …