An Anatomization for Classification Skin Lesion Using Custom CNN Framework

Abstract

A skin lesion is a growth or appearance anomaly in the dermis that is different from the rest of the body. A tiny scratch or something more serious like skin cancer might be completely unnoticeable. The use of deep learning (DL) architectures for image categorization in several fields, including dermatology, has shown impressive results. The expectations produced by DL for applications like image-based diagnostics have necessitated that non-experts become acquainted with algorithms’ working principles. Obtaining hands-on experience using these tools through a simplified however realistic method, in our opinion, may significantly aid in their intuitive comprehension. Even students without a strong mathematical background may understand concepts like convolution by seeing the outcomes of DL algorithms’ operations on dermatological images. Furthermore, the ability to fine-tune hyperparameters & even computer code expands the reach of basic intuitive understanding of these methods without having complex computational & theoretical skills. It is now feasible because of recent technology advancements that have lowered technical & technological obstacles associated with usage of these tools, creation them available to a larger population. So, we present a custom CNN for training a CNN on a dataset that contains skin lesions images linked by distinct skin cancer groups. Activity is open-source & doesn't need any additional software to be installed to use it. Additionally, we describe the algorithm and its functions in detail, starting with the creation of the basic computer code building blocks and leading the reader through the execution of real-world examples, comprising visualization & assessment of results. We get a training accuracy of 100 percent & a validation accuracy of 98.32 percent using this method.