Scholarlink

<p>Cardiovascular diseases (CVDs) remain a major cause of mortality worldwide, and electrocardiograms (ECGs) are an important non-invasive tool for diagnosis. However, many ECG records are still stored as scanned paper images containing grid interference, noise, and variable image quality, which complicates automated analysis. This study presents a framework for four-class ECG image classification directly from paper-based records, covering Normal, Myocardial Infarction (MI), Abnormal Heartbeat (HB), and Post-Myocardial Infarction (PMI). The proposed method combines a targeted preprocessing pipeline with a CBAM-enhanced ResNet-18 architecture. The preprocessing stages are designed to suppress non-diagnostic background interference while preserving clinically relevant waveform morphology, and the attention mechanism refines feature learning in both channel and spatial dimensions. Under a unified training strategy, the proposed model achieved a test accuracy of 0.9674 and a macro-F1 score of 0.9702 on the independent test set. In direct comparison under the same evaluation protocol, the custom CNN reference model achieved 0.9022 test accuracy and 0.8950 macro-F1, indicating clear gains from the proposed architecture. These findings suggest that the combination of tailored preprocessing and attention-enhanced residual learning is effective for paper-based ECG image classification and may provide useful support for automated analysis of legacy ECG archives and resource-limited clinical settings.</p>

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

CBAM-Enhanced ResNet for Robust Multi-Class ECG Image Classification

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月15日

查看全文
  • Yongting Liu,
  • Zixuan Ni,
  • Tongshuai Guo,
  • Meng Tang

摘要

Cardiovascular diseases (CVDs) remain a major cause of mortality worldwide, and electrocardiograms (ECGs) are an important non-invasive tool for diagnosis. However, many ECG records are still stored as scanned paper images containing grid interference, noise, and variable image quality, which complicates automated analysis. This study presents a framework for four-class ECG image classification directly from paper-based records, covering Normal, Myocardial Infarction (MI), Abnormal Heartbeat (HB), and Post-Myocardial Infarction (PMI). The proposed method combines a targeted preprocessing pipeline with a CBAM-enhanced ResNet-18 architecture. The preprocessing stages are designed to suppress non-diagnostic background interference while preserving clinically relevant waveform morphology, and the attention mechanism refines feature learning in both channel and spatial dimensions. Under a unified training strategy, the proposed model achieved a test accuracy of 0.9674 and a macro-F1 score of 0.9702 on the independent test set. In direct comparison under the same evaluation protocol, the custom CNN reference model achieved 0.9022 test accuracy and 0.8950 macro-F1, indicating clear gains from the proposed architecture. These findings suggest that the combination of tailored preprocessing and attention-enhanced residual learning is effective for paper-based ECG image classification and may provide useful support for automated analysis of legacy ECG archives and resource-limited clinical settings.