An Improved Skin Lesion Classification Using a Hybrid Approach with Active Contour Snake Model and Lightweight Attention-Guided Capsule Networks

<p dir="ltr">This article presents a novel deep learning approach for the <b>automated classification of skin lesions</b>, aiming to improve the <b>early and accurate diagnosis of skin cancer</b>. It highlights the challenges faced in clinical settings, such as long waiting times and the difficulty of distinguishing between benign and malignant lesions due to the complex visual characteristics of skin abnormalities.</p><p dir="ltr">Although deep learning models like Convolutional Neural Networks (CNNs) have shown promise, they still struggle with aspects like <b>precise lesion boundary detection</b>, <b>feature relationships</b>, and <b>contextual understanding</b>. To overcome these issues, the authors propose a hybrid model that integrates:</p><ul><li><b>Active Contour (AC) segmentation (snake models)</b> to enhance lesion boundary detection,</li><li><b>ResNet50</b> as a backbone for extracting features,</li><li><b>Capsule Networks</b> augmented with <b>lightweight attention mechanisms</b> to improve the discrimination and spatial understanding of features.</li></ul><p dir="ltr">The model is optimized using <b>Stochastic Gradient Descent (SGD)</b> and tested on two widely used skin lesion datasets—<b>HAM10000</b> and <b>ISIC 2020</b>. The results are impressive, showing:</p><ul><li><b>98% classification accuracy</b>,</li><li><b>97.3% AUC-ROC</b> score.</li></ul><p dir="ltr">These outcomes demonstrate the model's robustness and its potential to <b>outperform current state-of-the-art (SOTA) methods</b>, making it a valuable tool for assisting dermatologists in making faster and more accurate diagnoses.</p>