Project 02 — Computer Vision

Wound
Image
Classification

A deep learning image classification pipeline for automated wound condition recognition — from raw image datasets to a trained Keras model for healthcare monitoring.

TypeComputer Vision / CNN
FrameworkTensorFlow / Keras
ContextUTeM Research
TensorFlowKerasPython CNNImage ClassificationHealthcare AI
Main wound classification project demo

Architecture

Model Design

A convolutional neural network pipeline designed for multi-class wound image classification, covering the layers, backbone, and output classes used in the final model.

Input

Wound Images
(224 x 224)

→

CNN Layers

Conv + Pooling
Feature Maps

→

Classifier

Dense + Softmax
3 Classes

Architecture Notes

Full CNN model architecture and summary
Model Overview

Full Model Architecture

Full model diagram and summary used to document the final network layout.

CNN model parameters and layer configuration
Parameters

Layer Parameters

Layer parameters, feature-map sizes, and trainable parameter counts from the trained model.

Custom CNN layers designed for the wound classification model
Custom Design

Custom CNN Layers

Custom convolution blocks, pooling flow, and dense-layer design used in the final architecture.

The model combines convolution, pooling, and dense classification layers with TensorFlow/Keras training for three wound classes.

Dataset

Data & Preprocessing

Dataset details, augmentation strategies, and class distribution. Update the numbers below with your actual dataset stats.

4440 Total Images

All images were augmented

3 Classes

Healthy / Mild / Severe

95.54% Train / Val Split

Training accuracy achieved on the final run.

Preprocessing Image resizing, normalization, and channel standardization. Images were prepared for CNN training with consistent sizing, normalized inputs, and a standardized RGB pipeline.
Augmentation Data augmentation to improve generalization. Augmentation was used to expand the effective dataset and improve robustness across wound-image variations.
Class Balance Class balance management for wound categories. Dataset preparation focused on keeping the healthy, mild, and severe classes usable for stable CNN training.
Training Setup Google Colab with GPU acceleration. Training was run in TensorFlow/Keras with GPU support for faster iteration and model evaluation.

Visuals

Images & Results

Training curves, confusion-matrix output, and representative wound samples used to document model behavior and classification quality.

Training and validation curves Training & Validation Loss / Accuracy Curves
Confusion matrix Confusion Matrix
Healthy wound sample Sample Wound Images — Healthy
Mild wound sample Sample Wound Images — Mild

More Information

Additional Context

Additional background on the project motivation, clinical relevance, and research value behind the wound image classification workflow.

This project was developed to support tissue engineering research with a faster and more consistent wound assessment workflow.

3rd INOTEK 2024
Award Placement

Project Notes

Research support: This monitoring system helps streamline tissue engineering research by improving how wound conditions are documented and reviewed.

Department impact: The workflow can help accelerate scaffold development by giving the tissue engineering team faster access to organized image-based results.

Technical contribution: The project included dataset optimization for CNN training and a focused study on how convolution and dense layer design affected model performance.

Broader relevance: The work contributes to better healthcare research outcomes and aligns with SDG 3: Good Health and Well-Being.

← Previous Wound RF Classification All Projects Next → DorperNet IoT System