Modified Le-Net Model with Multiple Image Features for Skin Cancer Detection.

Abstract

Computer-based technologies significantly improve melanoma and non-melanoma skin cancer detection by providing non-invasive, cost-effective, and rapid diagnostic solutions. In this context, the study proposes a novel Deep Learning (DL)-based skin cancer detection approach that leverages an advanced segmentation technique called Improved DeepJoint Segmentation (IDJS). This method is designed to enhance the accuracy and precision of the detection process. Initially, the proposed Modified LeNet (MLeNet)-based model applies a Gaussian filter during preprocessing to reduce speckle noise in the input skin images effectively. Following this, the preprocessed images undergo the IDJS segmentation process, which effectively partitions the cancerous regions with high accuracy. Subsequently, three types of features are extracted from the segmented images and they are Multi-Texton Histogram (MTH)-based features, Improved Pyramid Histogram of Oriented Gradient (IPHOG)-based features, and Median Binary Pattern (MBP). These extracted features serve as the input to the MLeNet model for the final skin cancer detection. The datasets used in this work are the HAM10000 dataset and the ISIC 2019 dataset. With a positive metric value of 0.952, the MLeNet model outperforms the traditional models, with LeNet achieving the highest score of 0.932.