Performance Evaluation of Novel Convolution Neural Network Architecture for Melanoma Skin Cancer Diagnosis on Different Hardware Processing Units

Abstract

Skin cancer is one of the major public health concerns among the white population with more than a hundred thousand cases every year. Melanoma is one of the deadliest forms of skin cancer which is responsible for thousands of deaths in US alone in recent years and, therefore, early diagnosis is very important to increase the survival rate of melanoma patients. In last few years’ Deep neural networks have been utilized by researchers to build best models for classifying or diagnosing skin cancer. In this paper Deep neural network-based CNN architectures to classify Melanoma is proposed. The CNN architecture proposed in this work is implemented on CPU, GPU and TPU and the performance of the model is shown on all these platforms. The proposed model is compared to other works done so far for melanoma diagnosis in terms of various performance metrics like prediction accuracy, specifity, sensitivity and it is observed that the proposed models outperformed them. The dataset utilized in training and testing the proposed models is ISIC archive dataset which contains 4750 skin images for two classes i.e. melanoma and benign. The results of our study have proved that utilizing GPU and TPU speeds up the training 38 times faster than CPU and can accelerate the performance of CNN for features extraction, optimization and classification of skin cancer images and the proposed model has outperformed the other models compared with it.

Innovating Skin Cancer Detection with Transfer Learning, Feature Fusion, Efficient Data Management, Data Augmentation and Interdisciplinary Expertise for Better Outcomes

“…In our testing environment, the aforementioned techniques were implemented, and TPU processing units were utilized to optimize their performance and accelerate processing [25].…”