An Optimized Ensemble Framework for Multi-Label Classification on Long-Tailed Chest X-ray Data

Abstract

Chest X-rays (CXR) are essential in the diagnosis of lung disease, but CXR image classification is challenging because patients often have multiple diseases simultaneously. This requires multi-label classification to identify multiple abnormalities within a single image, which is complicated by different disease patterns and overlapping pathologies. In addition, CXR image classification faces the problem of long-tail distribution, with few common and mostly rare diseases, which can lead to biased predictions, especially for rare classes. There have been limited attempts to address these challenges in the medical domain, and applying general domain approaches to medical data may not be straightforward due to the unique characteristics of medical data. This paper presents an optimized ensemble framework to solve multi-label long-tailed classification on the MIMIC-CXR-LT dataset, which is the main objective of the ICCV CVAMD 2023 workshop competition, CXR-LT: Multi-Label Long-Tailed Classification on Chest X-Rays. Various experiments have been conducted, from architecture design to data augmentation, to identify the most suitable components. The proposed framework improves the performance of the long-tail distribution classification problem on class-imbalanced multi-label medical images and is placed in the top ranks in the CXR-LT competition.