Evaluating the Impact of Class Imbalance on Breast Ultrasound Image Classification

Class imbalance is a common challenge in classification problems in machine learning, where automated systems often struggle due to uneven class distributions. In medical imaging, class imbalance is directly related to epidemiological measures such as incidence. Although different strategies have been proposed to address class imbalance in the literature, there is a knowledge gap regarding the impact of disease incidence on the performance in medical image classification tasks. In this work, we aim to analyze the performance of imbalance compensation methods as a function of the distribution of classes in classification problems. For this purpose, we built a model to classify breast lesions and evaluated its performance in different imbalance scenarios. Three strategies were studied for imbalance compensation: focal loss, SMOTE and borderline-SMOTE. Focal loss consistently outperformed the baseline across imbalance levels ranging from 1:2 to 1:5, while bordeline-SMOTE showed advantages in more extreme imbalance scenarios of 1:8 or above. These findings provide guidelines for imbalance compensation methods in medical diagnosis tasks according to disease incidence.