Abstract
About this talk
In recent years, the growing availability of dermoscopic image datasets has supported the development of computer-aided diagnosis (CAD) methods to analyze and classify skin lesions. However, class imbalance remains a significant challenge in CAD development because diagnostic categories may contain substantially different numbers of images. This issue is especially relevant in multi-class dermatological datasets, where common lesion categories may have thousands of images, while less frequent categories have far fewer samples. For example, the ISIC 2019 dataset contains 25,331 images distributed in eight categories, with a markedly uneven class distribution. This imbalance can bias learning toward majority classes and reduce the representation of underrepresented lesion categories. This work reviews and analyzes class imbalance in skin lesion image classification from different conceptual perspectives, including rare cases, minority classes, small disjuncts, class distribution, intrinsic and extrinsic imbalance, relative imbalance, absolute rarity, and long-tailed learning. These concepts offer different perspectives on the origin, structure, and severity of imbalanced datasets. The review also examines learning strategies for addressing class imbalance, including data-level approaches such as oversampling, undersampling, and data augmentation, as well as algorithm-level approaches designed to reduce the influence of majority classes and improve the representation of minority categories. Particular attention is given to the distinction between conventional class imbalance and long-tailed distributions, in which a small number of head classes contain most of the samples. Understanding the different forms and characteristics of class imbalance is essential for interpreting CAD system performance and identifying limitations associated with underrepresented skin lesion categories. This review highlights the need to consider both the distributional characteristics of dermatological datasets and the learning strategies used to address them.
Speaker
About the speaker


