Diabetic Wound Analysis Application

A desktop application for analyzing diabetic wounds through image processing, segmentation, and machine learning techniques. The application provides automated wound assessment, color analysis, and generates detailed medical reports.

Table of Contents

1. Diabetic Wound Analysis Application
2. Features
3. Requirements
   • Dependencies
   • Model Files
4. Installation
5. Technical Details
   • Color Analysis
     • Basic Colors
     • Refined Color Ranges
   • Image Processing Pipeline
     • Preprocessing
     • Segmentation
     • Analysis
     • Classification
6. Command Line Usage
   • Command Line Arguments
7. File Structure
8. Output
9. Notes
10. Contributing

Features

• Interactive GUI built with Tkinter
• Image preprocessing and wound area detection
• Wound segmentation using U-Net and DeepSkin models
• Multi-color analysis with refined color ranges
• Automated wound clustering using UMAP and ResNet50
• AI-powered wound assessment report generation
• Step-by-step analysis workflow

Requirements

Dependencies

python >= 3.8
tkinter
PIL (Pillow)
opencv-python (cv2)
torch
transformers
numpy
scikit-learn
matplotlib
tensorflow
joblib
umap-learn

Model Files

The application requires pre-trained models in the following locations:

• Hackathon Official Data/Results/kmeans_model.joblib: Pre-trained K-means clustering model
• Hackathon Official Data/Results/umap_model.joblib: Pre-trained UMAP dimensionality reduction model
• U-Net model file (.pth format)

Installation

1. Clone the repository
2. Install the required dependencies:

pip install pillow opencv-python torch transformers scikit-learn matplotlib tensorflow joblib umap-learn

3. Ensure you have the required model files in the correct directory structure

Technical Details

Color Analysis

The application analyzes wounds using the following color ranges in HSV format:

Basic Colors

• Red: (0-10, 50-255, 50-255)
• Dark Red: (170-180, 50-255, 50-255)
• Yellow: (20-30, 100-255, 100-255)
• Black: (0-180, 0-255, 0-50)
• Pink: (160-170, 50-255, 50-255)
• White: (0-180, 0-30, 200-255)
• Brown: (10-20, 50-255, 50-200)
• Purple: (130-160, 50-255, 50-255)

Refined Color Ranges

• Light/Dark Yellow
• Light/Dark Brown
• Light/Dark Pink
• Light/Dark Purple

Image Processing Pipeline

1. Preprocessing:

   • Contour detection and cropping
   • Image normalization
   • ResNet50 feature extraction

2. Segmentation:

   • U-Net or DeepSkin model segmentation (Download checkpoints from here)
   • Mask generation and wound extraction
   • Contour drawing and dilation

3. Analysis:

   • Color percentage calculation
   • Multi-part segmentation
   • Histogram visualization

4. Classification:

   • Feature extraction using ResNet50
   • UMAP dimensionality reduction
   • K-means clustering prediction

Command Line Usage

The application can also be run from the command line:

python test.py --model MODEL.pth --input IMAGE_PATH [OPTIONS]

Command Line Arguments

• --model, -m: Model file path or "deepskin"
• --input, -i: Input image path(s)
• --mask-threshold, -t: Mask probability threshold (default: 0.5)
• --scale, -s: Input image scale factor (default: 0.5)
• --bilinear: Use bilinear upsampling
• --classes, -c: Number of classes (default: 2)

Output

The application generates:

1. Segmented wound images
2. Color distribution histograms
3. Multi-part segmentation visualization
4. Cluster prediction
5. Detailed wound assessment report

Notes

• The application uses ResNet50 for feature extraction
• UMAP is used for dimensionality reduction before clustering
• Color analysis uses HSV color space for better accuracy
• Both U-Net and DeepSkin models are supported for segmentation

Contributing

Feel free to submit issues, fork the repository, and create pull requests for any improvements.