A Constructive Auto-Encoder Based Convolutional Neural Network Model for Skin Lesions Prediction Using Compression Model

In medical imaging, skin lesion prediction and classification is highly crucial while predicting skin malignancy. Various prevailing deep learning-based CAD diagnosis approaches show poor performance. It is incredibly challenging to diagnose skin lesions with complex features like artefacts, boundary analysis, low contrast images with poor foreground and background images, and constraint training datasets. Also, it relies on the appropriate tuning of millions of parameters that causes poor generalization, overfitting, and massive consumption of computing resources. This research concentrates on modelling an efficient framework that performs some preliminary processes like pre-processing, segmentation and classification of skin lesions for automated prediction of skin lesions. The anticipated framework is composed of five stages: 1) pre-processing with adaptive median filtering; 2) segmentation with enhanced active contour-based approach; 3) ROI region compression with hybrid Lampel Ziv Wolch; 4) Non-ROI region compression with Fractal Image Compression (FIC) and 5) Classification with auto-encoder CNN. The proposed auto-encoder-based CNN model is designed with sub-networks connected in series, which is more efficient than conventional approaches. The simulation is done in the MATLAB 2020a environment. The performance of the anticipated model is compared with some standard methods like CNN and CNN-PSO. The predicted model shows a better trade-off than the prevailing approaches.