Abstract
Skin cancer is among the most prevalent types of malignancy all over the global and is strongly associated with the patient's prognosis and the accuracy of the initial diagnosis. Clinical examination of skin lesions is a key aspect that is important in the assessment of skin disease but comes with some drawbacks mainly with interpretational aspects, time-consuming and healthare expenditure. Skin cancer if detected early and treated in time can be controlled and its deadly impacts arrested completely. Algorithms applied in convolutional neural network (CNN) could lead to an enhanced speed of identifying and distinguishing a disease, which in turn leads to early detection and treatment. So as to eliminate these challenges, optimized CNN prediction models for cancer skin classification is studied in this researche. The objectives of this study were to develop reliable optimized CNN prediction models for skin cancer classification, to handle the severe class imbalance problem where skin cancer class was found to be much smaller than the healthy class. To evaluate model interpretability and to develop an end-to-end smart healthcare system using explainable AI (XAI) such as Grad-CAM and Grad-CAM++. In this researche new activation function namely NGNDG-AF was offered specifically to enhance the capabilities of network fitting and generalization ability, convergence rate and reduction in mathematical computational cost. A research used an optimized CNN and ResNet152V2 with the HAM10000 dataset to differentiate between the seven forms of skin cancer. Model training involved the use of two optimization functions (RMSprop and Adam) and NGNDG-AF activation functions. Cross validation technique the holdout validation is used to estimate of the model's generalization performance for unseed data. Optimized CNN is performing well as compare to ResNet152V2 for unseen data. The efficacy of the optimized CNN method with NGNDG-AF was examined by a comparative study wirh popular CNN with various activation functions shows that better performance of NGNDG-AF, achieving the classification accuracy rates that are as high as 99% in training and 98% in the validation. The recommended system also involves the integration of the smart healthcare application as a central component to give the doctors as well as the healthcare providers diagnosing and tools that would assist in the early detection of skin cancer hence leading to better outcomes of the treatment.