Unveiling hidden risks: A Holistically-Driven Weak Supervision framework for ultra-short-term ACS prediction using...

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Acute Coronary Syndrome (ACS) poses a significant health risk.
• Early prediction of ACS is crucial for timely intervention.
• Coronary CT Angiography (CCTA) provides detailed cardiac imaging but detecting subtle risk factors remains challenging.

Purpose of the Study:

• To develop and evaluate MH-STR, a novel end-to-end framework for predicting three-month ACS risk.
• To leverage hybrid attention mechanisms and convolutional networks for enhanced lesion pattern detection in CCTA images.
• To improve the accuracy and reliability of ACS risk stratification using deep learning.

Main Methods:

• Proposed MH-STR, an end-to-end framework integrating hybrid attention and convolutional neural networks.
• Implemented a stage-wise transfer learning strategy for feature distillation and knowledge transfer.
• Introduced a wavelet-based multi-scale fusion module to address feature scale mismatches in a dual-branch architecture.

Main Results:

• MH-STR achieved an Area Under the Curve (AUC) of 0.834.
• The model demonstrated an F1 score of 0.82 and a precision of 0.92.
• MH-STR outperformed existing methods in predicting ACS risk from CCTA images.

Conclusions:

• MH-STR shows significant potential for improving ACS risk prediction accuracy.
• The framework effectively captures subtle and irregular lesion patterns indicative of ACS risk.
• This approach offers a promising tool for enhancing early detection and management of ACS.