Unraveling the Impact of Class Imbalance on Deep-Learning Models for Medical Image Classification

Abstract

The field of image analysis with artificial intelligence has grown exponentially thanks to the development of neural networks. One of its most promising areas is medical diagnosis through lung X-rays, which are crucial for diseases like pneumonia, which can be mistaken for other conditions. Despite medical expertise, precise diagnosis is challenging, and this is where well-trained algorithms can assist. However, working with medical images presents challenges, especially when datasets are limited and unbalanced. Strategies to balance these classes have been explored, but understanding their local impact and how they affect model evaluation is still lacking. This work aims to analyze how a class imbalance in a dataset can significantly influence the informativeness of metrics used to evaluate predictions. It demonstrates that class separation in a dataset impacts trained models and is a strategy deserving more attention in future research. To achieve these goals, classification models using artificial and deep neural networks implemented in the R environment are developed. These models are trained using a set of publicly available images related to lung pathologies. All results are validated using metrics obtained from the confusion matrix to verify the impact of data imbalance on the performance of medical diagnostic models. The results raise questions about the procedures used to group classes in many studies, aiming to achieve class balance in imbalanced data and open new avenues for future research to investigate the impact of class separation in datasets with clinical pathologies.