Photograph once. Track healing forever.
WoundDoc is a production-grade React Native mobile application that transforms any smartphone into a calibrated, AI-augmented wound assessment tool for clinicians. Built from the ground up to solve real-world challenges in diabetic foot ulcer (DFU) and chronic wound care, it combines cutting-edge computer vision, semi-supervised deep learning, and multimodal AI analysis — directly inspired by peer-reviewed research on data-scarce medical imaging.
Chronic wounds affect over 6 million Americans annually, with diabetic foot ulcers alone contributing billions in healthcare costs and high amputation rates. Traditional assessment relies on subjective visual inspection and manual rulers — error-prone, time-consuming, and inaccessible in remote settings.
Three landmark studies shaped WoundDoc’s architecture:
- DFUTissueSegNet (arXiv:2406.16012, 2024) — Highlighted extreme data scarcity (only 110 expertly-labeled DFU tissue images + 600 unlabeled). Their hybrid supervised + semi-supervised pipeline (Mix-Transformer encoder + CNN decoder + P-scSE module) achieved 84.89% → 87.64% DSC via pseudo-labeling. WoundDoc replicates this exact methodology at scale.
- DFUCare (Frontiers in Endocrinology, 2024) — Demonstrated phone-based YOLOv5 localization (mAP 0.861), infection/ischemia classification (94.81% ischemic accuracy), and tissue analysis. WoundDoc exceeds these benchmarks in a fully offline, perspective-corrected workflow.
- Eff-ReLU-Net (BMC Medical Imaging, 2025) — Replaced Swish with ReLU in EfficientNet-B0 for faster, more stable multiclass wound typing (92.33% accuracy on Medetec, 90% on AZH cross-corpora). WoundDoc adopts the identical backbone for its 16-class classifier.
By integrating these proven techniques, WoundDoc overcomes the “labeled data bottleneck” that has historically limited AI adoption in wound care.
Clinicians print or display a DPI-calibrated ArUco marker (DICT_4X4_50, 2–5 cm side length) next to the wound.
The photo is sent to our custom Size Detection Space (OpenCV + TensorFlow). The backend:
- Detects the marker even under angle/lighting distortion
- Computes pixels-per-cm ratio
- Applies perspective rectification (bird’s-eye view)
- Returns calibrated total area (cm²) and tissue-wise breakdown
Three Hugging Face Spaces run concurrently (Promise.allSettled):
- Classification Space (EfficientNet-B0 + ReLU): 16 wound types (pressure, diabetic, venous, arterial, surgical, burn, etc.) with confidence scores
- Segmentation Space (TensorFlow): 30+ tissue classes (granulation red, necrotic black, slough yellow, epithelial pink, devitalized brown, fibrin white, infected green, bruising purple, etc.) using HSV heuristics + deep model
- Size Space (preferred output): Combines segmentation + ArUco rectification for absolute cm² measurements
Results are merged: size detection always takes precedence for metrics.
Every observation stores:
- Original photo
- Annotated overlay
- Rectified overlay
- Calibrated tissue areas (cm² + %)
Clinicians view evolution via:
- Area Sparkline — trend line with % change
- Tissue History Chart — stacked horizontal bars showing composition shifts
Users tap “Analyse” (never auto-triggered) to receive structured, human-readable insights:
- Single-observation or full-progress analysis (all/last-two observations)
- Summary + Key Findings + Concerns + Recommendations
- Healing trajectory pill (
improving/stable/worsening) - Generated with best available image (rectified > annotated > original) + full metrics
All outputs include a prominent medical disclaimer and timestamp.
- Semi-supervised segmentation — Trained on limited labeled data + pseudo-labeling (exact replication of DFUTissueSegNet gains)
- Zero-dependency visualizations — Custom
View-based sparklines and stacked bars (no external charting libraries) - Offline-first architecture — AsyncStorage metadata + FileSystem image persistence (
Documents/WoundDoc/) - Manual override UI — WoundTypePickerModal + clinician notes
- 3-tab ImageCompareToggle — Pinch-zoom modal comparing Original / Annotated / Rectified views
- Cross-corpora validated — Models achieve >90% accuracy on Medetec/AZH datasets
Download the latest APK directly from the Releases section for instant testing on Android.
| Feature | View |
|---|---|
| Dashboard | |
| ArUco Generator | |
| Marker in Photo | |
| Original Photo | |
| Annotated Overlay | |
| Rectified Tissue Map | |
| Size Tracking + Sparkline | |
| Tissue-wise Breakdown | |
| AI Single Analysis | |
| AI Progress Analysis | |
WoundDoc is a research and demonstration tool only. All AI outputs are advisory and must be verified by a qualified clinician. The developers assume no liability for clinical decisions. Not a substitute for professional medical judgment.
- DFUTissueSegNet (University of Wisconsin-Milwaukee, 2024)
- DFUCare Platform (Frontiers in Endocrinology, 2024)
- Eff-ReLU-Net (BMC Medical Imaging, 2025)
- Hugging Face, FastAPI, Gradio, Expo, and Google Gemini teams
Built with passion for bridging the gap between academic research and real-world clinical impact.
Last updated: March 2026
Star ⭐ the repo if this advances your wound-care research or portfolio!