Enhanced Skin Cancer Classification Through a Hybrid Optimized Approach: Deep Echo Network Machine Utilizing Pelican-Optimized Deep Kohonen Features

The global rise of skin cancer, accredited primarily to abnormal radiation-induced cellular growth, has resulted in millions of fatalities. Early detection and accurate diagnosis can potentially ameliorate the survival rate. However, traditional methods for skin lesion detection and classification (SLDC) employing machine learning (ML) and deep learning (DL) models have demonstrated limitations, notably higher stochastic gradient issues, increased mispredictions, elevated training losses, and decreased detection and classification accuracy. Addressing these limitations, we present an optimized Hybrid SLDC network (HOS-Net). The initial phase of the study saw the augmentation of the ISIC-2019 dataset, thereby increasing the quantity of images. These images were then preprocessed to normalize their size and data type. A Deep Convolutional Inverse Graphics Network (DCIGN) was subsequently developed to identify disease-affected regions from the preprocessed images. Following this, a Hybrid Deep Kohonen Network (HDKN) was introduced to extract disease-specific and disease-dependent features from the segmented images. Additionally, a Swarm-based Pelican Optimization Algorithm (SPOA) was implemented to extract the optimal features from the HDKN output features. Ultimately, a Deep Echo Network Machine (DENM) was utilized to classify various disease types using the pre-trained SPOA features. Simulations conducted on the ISIC-2019 dataset revealed that the proposed HOS-Net model achieved superior performance, with an accuracy of 99.13%, precision of 99.13%, recall of 99.25%, and an F1-score of 99.56%. This performance signifies the model's capability to accurately classify and capture positive cases in the data.