Advanced Augmentation and Ensemble Approaches for Classifying Long-Tailed Multi-Label Chest X-Rays

Abstract

Chest radiography is a common medical diagnostic procedure, often resulting in a long-tailed distribution of clinical findings. This challenges standard deep learning methods, which tend to favor more common classes and might miss less frequent but equally important "tail" classes. Chest X-ray diagnoses represent a multi-label problem due to the potential for multiple simultaneous diseases in patients. In this paper, we propose straightforward yet highly effective techniques to address the long-tailed imbalance in chest X-ray datasets. We specifically utilize EfficientNetV2 and ConvNeXt as our primary architectures, allowing the image sizes to influence architectural decisions. To counter dataset imbalance, we employ various basic and advanced augmentations. Mosaic augmentation is applied, and we alter the method of obtaining the label to manage this multilabel classification problem. We leverage the Binary Focal Cross-Entropy loss function and deploy several ensemble strategies to boost performance. These include Stratified K-Fold cross-validation and Test Time Augmentation. Our proposed method demonstrated its effectiveness during the Development and Testing phases of the CXR-LT: MultiLabel Long-Tailed Classification on Chest X-Rays competition. Our approach yields substantial results with an mAP of 0.354, securing a position within the top five.