Multilabel Chest X-Ray Image Classification via Category Disentangled Causal Learning

Qiang Li, Mengdi Liu, Rihao Chang, Weizhi Nie, Shaojin Bai, Anan Liu

Abstract: Chest X-rays (CXR) are widely used to diagnose chest diseases. Since patients often suffer from multiple diseases simultaneously, it is crucial to identify multiple abnormalities in a single CXR image, which is defined as a multilabel classification task. Recent methods aim to improve performance by leveraging label co-occurrences as prior knowledge. However, these statistical co-occurrences often introduce spurious correlations, which reduce the reliability of the model, and data imbalance further amplifies the harm of such spurious correlations for rare disease diagnosis. In this study, we proposed a category disentangled causal learning (CDCL) framework that considers both category-level and causal-level representations to provide robust and reliable CXR image diagnosis results. Specifically, we introduce the category attention (CA) mechanism to disentangle disease-specific features, enabling the model to effectively capture the discriminative features of each disease in the image. Additionally, we employ the label embeddings to learn a set of discriminative features at the global category level, complementing CA to enhance the effectiveness of category disentanglement. Causal intervention is then applied to the disentangled features to guide the model in learning true causal relationships, mitigating the impact of spurious correlations. The proposed CDCL framework was evaluated on the ChestX-Ray14 and CheXpert datasets, achieving mean AUC of 0.849 and 0.896, respectively. Ablation studies and visualization experiments demonstrated its competitiveness, particularly with significant improvements in rare disease identification.

External IDs:dblp:journals/tai/LiLCNBL26