Eff-ReLU-Net: a deep learning framework for multiclass wound classification

Chronic wounds have emerged as a significant medical challenge due to their adverse effects, including infections leading to amputations. Over the past few years, the prevalence of chronic wounds has grown, thus posing significant health hazards. It is now becoming necessary to automate the wound assessment mechanism to limit the dependence of healthcare practitioners on manual methods. Therefore, a need exists for developing an effective wound classifier that enables practitioners to classify wounds quickly and reliably. This work proposed Eff-ReLU-Net, an improved EfficientNet-B0-based deep learning model for accurately identifying multiple categories of wounds. More precisely, we introduced the ReLU activation function over the Swish in our Eff-ReLU-Net because of its simplicity, reliability, and efficiency. Additionally, we introduced three fully connected dense layers at the end to reliably capture more distinct features, leading to improved multi-class wound classification. We also employed augmentation approaches such as fixed-angle rotations at 90°, 180°, and 270°, rotational invariance, random rotation, and translation to improve data diversity and samples for better model generalization and combating overfitting. The proposed model's effectiveness is assessed utilizing the publicly available AZH and Medetec wound datasets. We also conducted the cross-corpora evaluation to show the generalizability of our method. The proposed model achieved an accuracy, precision, recall, and F1-score of 92.33%, 97.66%, 95.33%, and 96.48% on Medetec, respectively. However, for the AZH dataset, the attained accuracy, precision, recall, and F1-score are 90%, 89.45%, 92,19%, and 90.84%, respectively. These results validate the effectiveness of our proposed Eff-ReLU-Net method for classifying chronic wounds.