Practical fine-grained learning based anomaly classification for ECG image.

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Electrocardiography (ECG) is a vital sign for assessing heart function and diagnosing cardiovascular diseases.
• Automated ECG anomaly detection enhances diagnostic efficiency and reduces healthcare costs.
• ECG images are commonly stored in clinical institutions, necessitating effective image-based analysis.

Purpose of the Study:

• To develop a novel neural framework for automated anomaly detection directly from ECG images.
• To address the challenges of low resolution, high noise, and subtle signal differences in ECG images.
• To enable scalable and practical clinical applications of automated ECG analysis.

Main Methods:

• A novel neural framework designed for direct application to ECG images.
• Weakly supervised learning strategy requiring only image-level labeling.
• Classification of fine-grained ECG images to identify cardiology abnormalities.

Main Results:

• The proposed method effectively classifies fine-grained ECG images.
• Demonstrated effectiveness on two real-world ECG datasets.
• Significant savings in annotation time and cost due to the elimination of part annotations.

Conclusions:

• The novel neural framework enables effective automated detection of cardiology abnormalities from ECG images.
• Weakly supervised learning on ECG images offers a practical and cost-effective solution for clinical applications.
• This approach enhances the scalability and utility of automated ECG analysis in healthcare settings.