A Comparative Study on Skin Cancer Detection: Multi-Class Vs. Binary

Abstract

Skin cancer, particularly melanoma, is a major public health concern due to its high fatality rate. Early diagnosis is crucial for improving patient outcomes, and advances in computer-aided diagnostic systems based on deep learning have showed promise in increasing diagnostic accuracy. This study examines two methods for handling the multi-class classification issue in skin cancer diagnosis. The first strategy utilizes a single EfficientNet-b0 model to classify all classes at once, whereas the second approach, that can be thought as "waterfall" method, employs a sequence of binary classifiers, each designed to detect one specific class at a time. Both techniques were evaluated on the ISIC 2018 dataset, and the findings show that the waterfall like strategy improves classification accuracy by around 8%. This study illustrates the potential benefits of sequential binary classification in dealing with complicated multi-class problems in medical image analysis especially for skin cancer; nevertheless, more research with other metrics is required to corroborate these findings and explore different network models.