Prediction of Skin Lesions Using Integrated Multi-Layered Network Model with Baseline Learning Approaches

Abstract

Skin cancer has become more common in recent decades, raising severe concerns about world health. Creating an automated system to distinguish between benign and malignant images is challenging because of the subtle variations in how skin lesions appear. This study introduces Computer-Aided Diagnosis (CAD) system that offers high classification accuracy while maintaining low computing complexity for categorizing skin lesions. The system incorporates a pre-processing stage that uses morphological filtering to remove hair and artefacts. With the least minimum of human interaction, deep learning techniques are employed to separate skin lesions automatically. Image processing methods are currently being utilized to investigate the automated implementation of the prediction criteria for distinguishing between benign and malignant melanoma lesions. Various pre-trained convolutional neural networks (CNNs) with multi-layered (ML-CNN) are under examination for the classification of skin lesions as either benign or malignant. The best performance is achieved when RF, k-NN and XGBoost are combined, according to average 5-fold cross-validation findings. The outcomes also demonstrate that data augmentation works better than acquiring novel images for training and testing purposes. The experiment results show that the suggested diagnostic framework performs better than existing methods when used on actual clinical skin lesions, with accuracy at 97.5%, F1-score at 91.3%, precision at 96.5%, sensitivity at 89.2% and specificity at 96.7%. It also takes 2.6 seconds to complete with the MNIST dataset and accuracy at 98.2%, F1-score at 92.5%, precision at 98.4%, sensitivity at 92.3% and specificity of 97.2% with the ISIC dataset. This indicates that medical professionals can benefit from using the suggested framework to classify various skin lesions.