An AI-powered wound assessment system designed to analyze wound images, measure wound dimensions, track healing progress over time, and generate clinically useful reports.
Project Status: 🚧 Initial Planning / Architecture Phase Implementation Status: No implementation has been completed yet.
Wound healing is commonly monitored by visually inspecting the wound and manually estimating its size and progress. This process can be subjective, time-consuming, and difficult to compare consistently across multiple assessments.
This project aims to develop an AI-assisted system that can analyze wound images and provide objective measurements and progress information.
The proposed system will:
- Detect the wound region in an image
- Precisely segment the wound boundary
- Calculate wound area and other measurements
- Compare wound assessments across time
- Estimate healing progress
- Provide a recovery timeline
- Generate an automated clinical report
- Allow a clinician to review and verify AI-generated results
The project is intended as a clinical decision-support system, not as an autonomous diagnostic system.
Doctors and healthcare professionals often estimate wound size and healing progress manually.
This can introduce:
- Subjectivity in measurements
- Variation between different assessments
- Difficulty comparing images taken at different times
- Additional workload for clinicians
- Limited quantitative information about healing trends
The goal of this project is to develop an AI-assisted workflow that converts wound images into measurable and trackable healing information.
The primary objectives of the project are:
- Capture and upload wound images.
- Detect the wound region using deep learning.
- Segment the wound at pixel level.
- Calculate wound dimensions and area.
- Compare current and previous wound assessments.
- Estimate healing percentage and healing trends.
- Provide a recovery timeline.
- Generate an automated clinical report.
- Allow clinicians to verify or adjust AI-generated measurements.
- Maintain historical assessment data for each wound.
The core AI pipeline is planned as:
Wound Image
│
▼
┌─────────────┐
│ YOLO │
│ Detection │
└──────┬──────┘
│
▼
┌─────────────┐
│ U-Net │
│ Segmentation│
└──────┬──────┘
│
▼
┌─────────────┐
│ OpenCV │
│ Measurement │
└──────┬──────┘
│
▼
Wound Measurements
│
▼
Healing Progress Analysis
│
▼
Clinical Report
YOLO is planned for wound detection.
Its purpose is to identify the approximate location of the wound within the image before performing detailed segmentation.
U-Net is planned for pixel-level wound segmentation.
Instead of only producing a bounding box, U-Net will be used to identify the precise wound boundary.
OpenCV will be used for image processing and measurement operations, including:
- Image preprocessing
- Cropping
- Contrast adjustment
- Mask processing
- Contour extraction
- Pixel-area calculation
- Geometric measurements
- Calibration
The proposed YOLO → U-Net → OpenCV pipeline is described in the project architecture document.
A major consideration of this project is converting image pixels into meaningful physical measurements.
A segmentation model can determine the number of wound pixels, but pixel measurements alone do not directly represent real-world dimensions such as cm or cm².
Therefore, the proposed system will investigate:
- A physical reference marker/ruler in the image
- A controlled image-capture setup
- Consistent camera distance and angle
- Lighting and image-quality requirements
The calibration strategy will be finalized before implementing clinical area measurements.
Important: Reporting a measurement such as
4.2 cm²without a reliable physical reference would create a false sense of precision.
The architecture proposal specifically identifies calibration as an important design decision that should be addressed before measurement implementation.
The planned system will be divided into separate application and machine-learning services.
┌──────────────────────┐
│ React + TS │
│ Clinical Dashboard │
└──────────┬───────────┘
│
▼
┌──────────────────────┐
│ Node.js + Express │
│ Main Backend API │
└───────┬───────┬──────┘
│ │
┌──────────┘ └───────────┐
▼ ▼
┌─────────────────┐ ┌─────────────────┐
│ PostgreSQL │ │ MinIO / S3 │
│ Clinical Data │ │ Image Storage │
└─────────────────┘ └─────────────────┘
│
▼
┌──────────────────────┐
│ FastAPI │
│ ML Service │
└──────────┬───────────┘
│
▼
┌──────────────────────┐
│ YOLO → U-Net → │
│ OpenCV │
└──────────┬───────────┘
│
▼
┌──────────────────────┐
│ Measurements + AI │
│ Results + Confidence │
└──────────────────────┘
The architecture separates the application/clinical workflow from the machine-learning workload. Node.js/TypeScript is planned for the application layer, while FastAPI/Python will handle the AI pipeline.
| Layer | Technology | Planned Purpose |
|---|---|---|
| Frontend | React + TypeScript | Clinical dashboard and image comparison interface |
| Main Backend | Node.js + Express + TypeScript | Application and clinical workflow |
| ORM | Prisma | Database access and type-safe queries |
| Database | PostgreSQL | Patients, wounds, assessments and measurements |
| Image Storage | MinIO / S3 | Wound image storage |
| ML API | FastAPI | AI inference service |
| Object Detection | YOLO | Wound localization |
| Segmentation | U-Net | Precise wound boundary segmentation |
| Image Processing | OpenCV | Preprocessing and measurements |
| Cache | Redis | Optional caching/rate limiting |
| Async Processing | RabbitMQ | Optional future asynchronous inference |
| Interoperability | FHIR | Future healthcare-system integration |
| Deployment | Docker | Containerized development and deployment |
These technologies are proposed, not yet implemented. The architecture document describes PostgreSQL, MinIO, FastAPI, YOLO, U-Net and OpenCV as the core planned components, with Redis and RabbitMQ as later/optional additions.
The clinical data is expected to follow a relational structure:
Patient
│
├── Wound
│ │
│ ├── Assessment
│ │ │
│ │ ├── Image
│ │ ├── AI Result
│ │ ├── Measurements
│ │ └── Clinician Review
│ │
│ └── Assessment
│
└── Wound
- Patient
- Wound
- Assessment
- Image
- Measurement
- AI Result
- Clinician Review
- Report
PostgreSQL is proposed because the project contains naturally relational data such as patient → wound → assessment → measurement, with relationships and constraints that should be enforced by the database.
Wound images are planned to be stored separately from clinical metadata.
PostgreSQL
│
├── Patient information
├── Wound information
├── Assessment information
└── Image object key / metadata
MinIO / S3
│
├── wound-image-001
├── wound-image-002
└── wound-image-003
The database will store metadata and references to images, while the actual image files will reside in object storage.
For local development, MinIO is proposed, with an S3-compatible storage system planned for deployment.
This keeps large binary image data separate from structured clinical data.
The intended workflow is:
1. Clinician uploads wound image
↓
2. Backend validates the image
↓
3. Image stored in MinIO / S3
↓
4. Assessment record created
↓
5. Node.js sends image for AI analysis
↓
6. FastAPI receives analysis request
↓
7. YOLO detects wound
↓
8. U-Net segments wound
↓
9. OpenCV calculates measurements
↓
10. AI result stored
↓
11. Clinician reviews result
↓
12. Current assessment compared
with previous assessments
↓
13. Healing trend calculated
↓
14. Clinical report generated
This workflow follows the proposed upload and retrieval flow in the architecture document.
The system is planned to produce the following outputs:
A segmentation mask showing the detected wound region.
Original Image
+
Segmentation Mask
↓
Detected Wound Boundary
Estimated wound area based on the segmented region and the selected calibration method.
A comparison between previous and current wound measurements.
For example:
Previous Area : 10.0 cm²
Current Area : 7.5 cm²
Estimated reduction = 25%
The exact healing metric and formula will be defined during implementation and validation.
The system will maintain historical assessments to visualize wound progression over time.
Day 0 → 10.0 cm²
Day 7 → 8.5 cm²
Day 14 → 7.1 cm²
Day 21 → 5.8 cm²
↓
Healing Trend
A future report may contain:
- Patient/wound information
- Assessment date
- Original wound image
- Segmentation result
- Wound measurements
- Previous measurements
- Healing trend
- AI confidence
- Model version
- Clinician verification
- Observations
The AI output will not automatically be treated as the final clinical measurement.
The proposed workflow allows a clinician to:
- Review the AI-generated wound boundary.
- Accept the result.
- Adjust the boundary if required.
- Verify the measurement.
- Store both the original AI result and the clinician-verified result.
This distinction is important because it allows the project to evaluate how closely the model agrees with clinical assessment over time.
Since the project deals with potentially sensitive medical information, security will be considered from the beginning.
Planned considerations include:
- Authentication
- Role-based access control
- Audit logging
- HTTPS/TLS
- Protected image access
- Upload validation
- Consent tracking
- Model version tracking
These are planned architectural requirements and are not currently implemented.
FHIR is planned as a future interoperability layer.
Potential FHIR resources include:
PatientObservationDiagnosticReportDocumentReference
FHIR would allow the system to exchange relevant information with external healthcare systems without replacing the project's internal REST API.
The project will be developed incrementally.
- Design PostgreSQL schema
- Set up Node.js/Express backend
- Set up React frontend
- Implement Patient → Wound → Assessment structure
- Implement secure image upload
- Set up MinIO object storage
- Set up FastAPI ML service
- Implement image preprocessing
- Prepare wound detection dataset
- Train/evaluate YOLO model
- Prepare wound segmentation dataset
- Train/evaluate U-Net model
- Implement wound segmentation
- Implement OpenCV measurement pipeline
- Define and implement calibration
- Store AI results
- Build wound history/timeline
- Compare previous and current images
- Compare segmentation masks
- Implement clinician verification
- Calculate healing percentage/trends
- Generate clinical reports
- Implement RBAC
- Implement audit logging
- Protect image access
- Add model versioning
- Explore FHIR integration
- Redis caching
- RabbitMQ asynchronous inference
- Application/ML monitoring
- Cloud deployment
The phase structure follows the proposed build order in the architecture document.
The repository structure is expected to evolve during implementation. An initial target structure is:
wound-healing-analysis/
│
├── frontend/
│ └── React + TypeScript application
│
├── backend/
│ └── Node.js + Express + TypeScript
│
├── ml-service/
│ └── FastAPI + Python
│
├── ml/
│ ├── yolo/
│ ├── unet/
│ ├── preprocessing/
│ └── measurement/
│
├── data/
│ └── Dataset documentation
│
├── docs/
│ ├── architecture/
│ └── research/
│
├── docker/
│
├── README.md
└── .gitignore
This is a planned structure. Directories will be added as implementation begins.
Several challenges will need to be addressed during development.
A suitable wound-image dataset with reliable annotations will be required for training and evaluation.
The wound boundary may be difficult to identify because of:
- Different wound types
- Lighting variations
- Skin-tone variation
- Shadows
- Blood or exudate
- Irregular wound boundaries
- Background objects
Pixel measurements cannot automatically be interpreted as real-world measurements without a reliable scale.
Healing comparisons become more meaningful when images are captured under reasonably consistent conditions.
AI-generated measurements should be evaluated against appropriate ground-truth or clinician-verified measurements.
Every AI-generated assessment should eventually record which model/version produced the result so that historical results remain traceable.
The architecture proposal specifically highlights calibration, clinician verification, model versioning, and consistent capture conditions as important considerations.
At the current stage:
- No AI model has been implemented.
- No YOLO model has been trained.
- No U-Net model has been trained.
- No image segmentation pipeline exists yet.
- No wound measurement pipeline exists yet.
- No healing prediction model exists yet.
- No frontend has been implemented.
- No backend has been implemented.
- No database has been implemented.
- No clinical report generator has been implemented.
This repository currently represents the planning and architecture stage of the project.
Possible future improvements include:
- Improved wound segmentation models
- More accurate wound-area estimation
- Multi-class wound analysis
- Wound severity assessment
- Tissue classification
- Healing-rate prediction
- Personalized recovery estimates
- Automated longitudinal trend analysis
- Model confidence visualization
- Explainable AI features
- FHIR-based interoperability
- Cloud deployment
- Monitoring and model performance tracking
Model evaluation will be introduced once the dataset and models are available.
Potential evaluation metrics include:
- Precision
- Recall
- mAP
- Dice Score
- Intersection over Union (IoU)
- Precision
- Recall
- Mean Absolute Error
- Relative measurement error
The final evaluation methodology will depend on the available longitudinal dataset and clinically validated ground truth.
This project is being developed as an academic/research prototype and clinical decision-support concept.
It is not intended to replace professional medical judgment, diagnosis, or treatment decisions.
AI-generated measurements and predictions should be reviewed by a qualified healthcare professional before being used for clinical decision-making.
The repository will gradually include documentation covering:
- System architecture
- Dataset selection
- Data preprocessing
- YOLO training
- U-Net training
- Image segmentation
- Wound measurement
- Calibration
- Healing-progress calculation
- Backend API
- Database design
- Clinical workflow
- Model evaluation
- Deployment
Current Stage: 🟡 Planning / Architecture
Architecture ██████████ Planned
Backend ░░░░░░░░░░ Not Started
Frontend ░░░░░░░░░░ Not Started
Dataset ░░░░░░░░░░ To Be Finalized
YOLO ░░░░░░░░░░ Not Started
U-Net ░░░░░░░░░░ Not Started
OpenCV Pipeline ░░░░░░░░░░ Not Started
Measurement ░░░░░░░░░░ Not Started
Healing Analysis ░░░░░░░░░░ Not Started
Clinical Report ░░░░░░░░░░ Not Started
AI-Based Wound Healing Progress Analysis
The project aims to combine computer vision, deep learning, backend engineering, and healthcare-oriented software design into an end-to-end wound assessment platform.
The long-term goal is to build a system where a clinician can upload a wound image and receive a structured assessment containing:
Wound Image
↓
Wound Detection
↓
Precise Segmentation
↓
Measurements
↓
Historical Comparison
↓
Healing Trend
↓
Clinician Verification
↓
Clinical Report
The focus is not simply to build an image-classification model, but to develop a complete and traceable workflow around wound measurement and longitudinal healing assessment.