SGMoE: Semantics-Guided Mixture of Experts for Multi-class Disease Classification Based on Radiology Reports

ACL ARR 2025 February Submission3918 Authors

Abstract: Radiology reports contain detailed image description that is crucial for clinicians in decision-making, and automated disease classification based on radiology reports can be more effective than image-based classification. Although deep learning achieves promising performance on this task, existing approaches struggle with increasing class complexity when multi-class disease classification is considered, where multiple distinct modes can coexist within the feature distribution of a single class, leading to highly complicated decision hypersurfaces. To address this challenge, we propose Semantics-Guided Mixture of Experts (SGMoE) for report-based multi-class disease classification. SGMoE specializes multiple classification experts in handling different disease modes, each learning class boundaries within a specific subspace of the feature distribution. To guide the subspace allocation for each expert, SGMoE uses the report semantics to determine the expert assignment. This is achieved by clustering the report semantic embedding, and then an expert is assigned to determine specific classes in a certain cluster or clusters. Moreover, a gating network is designed to adaptively select appropriate experts for final classification, with a gating loss penalizing gating that contradicts with the expert assignment for model training. Experiments on an in-house dataset of 11,864 reports and the public CT-RATE dataset show that SGMoE achieves more accurate multi-class disease classification than existing text classification approaches.

Paper Type: Long

Research Area: Special Theme (conference specific)

Research Area Keywords: NLP Applications, Semantics: Lexical and Sentence-Level

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Data analysis

Languages Studied: English, Chinese

Submission Number: 3918