Advanced skin lesion segmentation and classification using adaptive contextual GLCM and deep learning hybrid models

Abstract

One of the most common and dangerous types of cancer, skin cancer, especially melanoma, is distinguished by its quick spread and high death rate if left untreated. It is impossible to overestimate the significance of early detection and precise diagnosis as they are essential for successful treatment and greatly raise patient survival rates. To improve skin lesion segmentation and classification, this work presents a state-of-the-art image analysis technique that uses hybrid deep learning models in conjunction with the Adaptive Contextual Gray Level Co-occurrence Matrix (GLCM). DeepLabV3 + is used to segment the skin lesion from pre-processed input images with high accuracy. A Long Short-Term Memory (LSTM) network trained on characteristics collected by the Adaptive Contextual GLCM is included to further evaluate the segmented images. Because of this special combination, the model can accurately capture contextual information and tiny texture differences, both of which are essential for differentiating between benign and malignant skin lesions. Our suggested approach performs very well in tasks involving both segmentation and classification. In particular, it attains a remarkable 98.34 % accuracy, 99.13 % specificity, 97.25 % sensitivity, 97.15 % Dice coefficient, and 98.6 % F1 score. These findings demonstrate the method’s stability and dependability, which makes it a potentially useful instrument for raising the detection accuracy of skin cancer. This approach has the potential to transform early detection procedures by providing improved diagnostic accuracy and resilience, eventually contributing to better patient outcomes and more effective management of skin cancer cases.