Enhancing Multi-Label Long-Tailed Classification on Chest X-Rays through ML-GCN Augmentation
HyeRyeong Seo, M.H. Lee, W. Cheong, HyeKyung Yoon, SoHyung Kim, Myungjoo Kang
Abstract
The classification of multi-label thoracic images presents a considerable challenge due to the severe intrinsic imbalances inherent in the dataset. During the testing phase, the model encounters both predominant (head) and less frequent (tail) classes, demanding not only proficiency in image feature extraction but also a comprehensive understanding of label relationships. Traditional medical image classifiers have historically relied on exploiting a small number of dominant head classes. Nevertheless, this approach often yields suboptimal classification outcomes. To resolve this issue, we propose an enhanced version of the Multi-Label Graph Convolutional Network (ML-GCN). Our approach integrates the incorporation of experts, each focusing on distinct aspects of the input dataset, class-balanced sampling, Log-Sum-Pooling (LSE pooling), an attention layer, and regularization through KL divergence. By synergistically applying these techniques, our model significantly outperforms the baseline vanilla ML-GCN, capitalizing on nuanced architectural adjustments. Through this comprehensive approach, we effectively demonstrate the versatility of our model in addressing the specific task of multi-label long-tailed classification within the realm of chest X-ray datasets. Furthermore, our methodology exhibits promising potential for extension to a diverse array of datasets characterized by long-tailed distributions, establishing a strong foundation for its application within various domains. In order to ensure the reproducibility of this study, we will make the source code publicly available: github.com/lisaseo9704/2023-ICCVW-CVAMD-NCIA500