A Deep Multiclass Learning Approach for Analysing Human Skin Lesions Using Ensemble Regional Dense Neural Networks (ERDNNs)

Abstract

Melanoma is considered to be one of the deadliest skin cancer types, which is found among 5% of western countries population. The lack of awareness and few practical difficulties in identification of the disease are the main factors in increasing the mortality rate due to skin cancer. The treatment given for this particular type of cancer remains problematic, because of its pigmentation resemblances with normal skin lesions. The clinical examinations of the Melanoma are mostly happening after the second stage of the lesions, which remains a huge problem in giving perfect treatment to the skin cancer within short duration. The Deep learning researchers remunerated more attention in implementing the procedure of Melanoma identification in computer vision-based approach to solve the inefficiency of manual inspection. The classification and prediction stages in Deep Learning are more accurate and time restraint factors for decreasing the mortality rate due to Melanoma. The research work carried out in gives more attention in giving the best rule setup in digital image processing as well as in deep learning procedure. The proposed Ensemble Regional Dense Neural Network (ERDNN) eradicates many issues faced previously during manual inspections and segments the affected parts accurately. The initial stage of the framework follows the basic digital image processing stages for picking the colour channels and enhancing the quality of the collected images from HAM10000 and HAM datasets. The feature, wrapper and embedded based feature selection process is carried out with convolutional method as well as with Deep Neural Network CReLu for Segmentation procedure. The furthermost procedure of the research is to examine the segmented images with proposed ERDNN for identifying the severity level of Melanoma. The comprehensive measurement is carried out with confusion metrices for testing the accuracy of the proposed ERDNN with other existing classification models. The performance of the ERDNN in concern to accuracy is 95% which is far better than other existing models.