Multiclass wound image classification using an ensemble deep CNN-based classifier.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Computational Biology

Background:

• Wound classification is crucial for effective treatment, yet manual methods are time-consuming and costly.
• Existing machine learning and deep learning approaches for wound classification have limitations.
• Accurate wound diagnosis supports optimal treatment strategies and improves patient outcomes.

Purpose of the Study:

• To develop a high-performance, ensemble Deep Convolutional Neural Network (DCNN) classifier for accurate wound image categorization.
• To improve the efficiency and reduce the cost of wound diagnosis for healthcare professionals.
• To create a robust decision support system for classifying various wound types.

Main Methods:

• An ensemble DCNN classifier was developed, integrating patch-wise and image-wise classification outputs.
• A Multilayer Perceptron (MLP) was used to combine the scores from the DCNN classifiers for superior performance.
• A 5-fold cross-validation strategy was employed for rigorous evaluation of the proposed method.

Main Results:

• The proposed classifier achieved maximum accuracies of 96.4% (binary) and 91.9% (3-class) with 5-fold cross-validation.
• Average accuracies were 94.28% (binary) and 87.7% (3-class) in the 5-fold cross-validation.
• On the Medetec dataset, accuracies of 91.2% (binary) and 82.9% (3-class) were achieved, outperforming common deep classifiers.

Conclusions:

• The developed ensemble DCNN classifier demonstrates significant potential for accurate wound image classification.
• The system can serve as an effective decision support tool, reducing financial and time costs in wound diagnosis.
• This AI-driven approach has broad applicability in clinical settings and wound management.