Feature Concatenation-Based Deep Learning Method for Multiclass Wound Classification

Chronic wounds represent a significant global healthcare challenge, affecting millions of individuals and leading to substantial financial burdens. Accurate classification of chronic wounds is crucial for guiding effective treatment strategies, improving patient outcomes, and reducing healthcare costs. However, existing methods in chronic wound classification predominantly focus on binary classification, with limited success in multi-class scenarios. Consequently, there is a need for diagnostic solutions that can reliably classify multiple wound types. To overcome these challenges, we present a novel feature concatenation-based deep learning method for reliable multiclass classification of chronic wounds. Specifically, we employed the DenseNet121 and XceptionNet models for generating effective and compact feature maps that are then passed to a Multilayer perceptron Classifier to accurately identify different types of wounds. The ability of DenseNet121 to truncate redundant features and XceptionNet to handle complex features with fewer computational resources allows us to generate compact and efficient feature maps. We evaluated the performance of our method using the publicly available AZH and Medetec datasets, as well as their augmented versions. Multiple augmentation techniques were employed to increase sample size and diversity, addressing overfitting and class imbalance issues. A rigorous experimental study performed on the AZH, Medetec, and augmented collections demonstrates the effectiveness of our method for multiclass wound classification.