Optimizing Wound Classification: A Deep Learning Method with VGG16 and Feature Selection for Clinical Application

The classification of chronic wounds is an important stage to ensure effective clinical management and treatment planning.This work proposes a Deep-Learning-based method to classify various wound types such as Diabetic, Pressure, Surgical, and Venous ulcers by using the VGG16 and VGG19 transfer learning models.The performance of both these fine-tuned models against the wound image dataset was compared.VGG16 performed better compared to VGG19 in multi-class and binary classification tasks, with better accuracy and generalization.The features were extracted using VGG16, followed by forward feature selection with the Fisher Discriminant Ratio, on the binary classification for 10 wound class pairs.The selected features were classified by SVM with a linear kernel for reducing bias and variance, hence making the model more robust.On binary classification, accuracy results were quite high, especially in differentiating normal skin from ulcerative conditions, reaching up to 99%.Noticeably, more challenging classifications, such as Diabetic vs Surgical and Surgical vs Venous, improved accuracy by over 7%, thus showing the effectiveness of feature selection in improving the model's performance.Our findings underline the effectiveness of transfer learning for medical image classification even in scenarios where limited data is available.The proposed method offers an automated solution for wound classification, a very useful tool for clinicians in the diagnosis and treatment of chronic wounds.