Scholarlink

We propose an early-fusion framework that combines deep features from chest X-rays (CXR) with bag-of-words (BoW) representations of pre-imaging clinical notes. Using eight CNN backbones to extract image features and BoW vectors from notes, our model learns adaptive modality weights via a single-layer attention module and feeds the fused embedding into XGBoost classifiers. On a real-world Vietnamese hospital dataset of 15416 paired CXRs and notes spanning seven pulmonary diseases, fusion improves accuracy and F1 score over image-only baselines across all backbones. Results indicate that even simple clinical text cues, when fused early with radiographic features, enhance robustness on minority classes and better reflect clinical reasoning.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Adaptive Early Fusion of Clinical Bag-of-Words and Chest X-Ray Deep Features for Pulmonary Disease Diagnosis

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Phuoc-Hai Huynh,
  • Thi-Diem Truong

摘要

We propose an early-fusion framework that combines deep features from chest X-rays (CXR) with bag-of-words (BoW) representations of pre-imaging clinical notes. Using eight CNN backbones to extract image features and BoW vectors from notes, our model learns adaptive modality weights via a single-layer attention module and feeds the fused embedding into XGBoost classifiers. On a real-world Vietnamese hospital dataset of 15416 paired CXRs and notes spanning seven pulmonary diseases, fusion improves accuracy and F1 score over image-only baselines across all backbones. Results indicate that even simple clinical text cues, when fused early with radiographic features, enhance robustness on minority classes and better reflect clinical reasoning.