Skin Cancer Detection Using Deep Learning Algorithms

Abstract

Skin disorders are very difficult to diagnose since many conditions have similar appearances, which makes it hard to tell them apart. While melanoma remains the most prevalent form of skin cancer, other illnesses are now known to be fatal. The most significant barrier to the development of automated dermatological systems is the scarcity of thorough and thorough data.This thesis introduces a robust deep learning architecture tailored for skin cancer diagnosis through the use of transfer learning on two types of convolutional neural networks (CNNs). These consist of a basic classifier and a two-tiered hierarchical classifier that makes use of two advanced CNNs. Differentiating between seven categories of nevi is the goal because precise diagnosis and treatment planning depend on it. The study's primary dataset is the HAM10000 dataset, a sizable collection of dermoscopic photos. Model performance is improved through the process of integrating multiple data sources. The outcomes unequivocally show how successful the DenseNet201 network is in this situation. This special combination reduces false positives while improving classification and Fmeasure accuracy. The test revealed that, surprisingly, the simple model performed better than the hierarchical two level model. The hierarchical approach is more straightforward than that, despite its attempt to provide classification at various levels. Specifically, the most successful level of binary classification is the first one, especially when it comes to differentiating lesions with and without nevus. This paper emphasizes the significance of applying deep learning methods particularly transfer learning to address the challenging skin cancer categorization issue. It is stressed how crucial data sets like HAM10000 are to the development of dermatological research. The outcomes validate the effectiveness of DenseNet201 in categorizing skin conditions and emphasize the necessity of refining the classification algorithm to produce more dependable, precise, and enhanced diagnoses, hence enhancing dermatological care.