Enhanced Lightweight Deep Learning Framework with Knowledge Distillation and Binary Whale Optimization for Diabetic Foot Ulcer Classification

Abstract

A common hopeless significance of diabetes is Diabetic Foot Ulcers (DFU), which require speedy and accurate analysis to avoid amputation and lower death. Computerized DFU identification has likely ended due to the rapid growth of deep learning; the computational requirements of existing high-performing models limit their application in quantifiable settings with limited resources. To overcome this challenge, this research focuses on addressing a key question: whether a lightweight DFU classification model can be strengthened using knowledge distillation and automated hyperparameter tuning to uphold the performance of the model that is suitable for edge development. This research presents an original, lightweight classification framework that combines the Binary Whale Optimization Algorithm (BWOA) with Knowledge Distillation (KD) to yield an effective and specialized DFU classification system. This technique creates soft probability labels by using a pretrained InceptionV3 model as a teacher. These are motivated by DFU-LWNet, a small convolutional neural network with little parameter overhead, which is a customizable student network. The baseline DFU-LWNet with KD imitates earlier best findings (96.23% accuracy) on experiments with the publicly available DFU Patch Dataset. The proposed model DFU-LWNet-BWOA achieves a significant accuracy gain of 96.8% compared to the previous models, and it also ensures real-time compatibility for the mobile applications by making the consistency of parameter count as 0.5M. This study mainly focuses on the deployable, intelligent, and scalable solution for the DFU model in an experimental setup and ensures the interaction between the model, knowledge distillation, and BWOA optimization.