Multi-instance learning with attention mechanism for coronary artery stenosis detection on coronary computed tomography angiography.

Abstract

Aims: Accurate detection of coronary artery stenosis (CAS) on coronary computed tomography angiography is vital for saving lives, as timely diagnosis can prevent severe cardiac events. However, this task remains challenging due to data complexity and variability in imaging protocols. Deep learning offers promising potential to automate detection, but robust methods are essential to address real-world challenges effectively and enhance patient outcomes.

Methods and results: A total of 900 cases with curved multiplanar reformations, pre-generated during routine clinical workflows, were used to train a multi-instance learning (MIL) model for detecting significant CAS (≥50% luminal obstruction) in the left anterior descending (LAD), right coronary artery (RCA), and left circumflex (LCX), comprising 776 LAD, 694 RCA, and 600 LCX reconstructions. Patient-level predictions utilized attention scores to quantify each slice's contribution, ensuring a robust and interpretable diagnostic approach. The model achieved the best performance for LAD [area under the curve (AUC): 0.92, 95% confidence interval (CI): 0.87-0.96; Brier score: 0.11], followed by RCA (AUC: 0.91, 95% CI: 0.82-0.999; Brier score: 0.09) and LCX (AUC: 0.92, 95% CI: 0.84-0.99; Brier score: 0.07). Calibration was good for LAD but less precise for RCA and LCX. Attention scores enhanced diagnostic precision by focusing on the most relevant slices.

Conclusion: This study highlights the potential of MIL models for CAS detection, with remarkable performance in the LAD. By using attention scores, the model effectively identifies key slices from real-world data, seamlessly integrating with routine clinical workflows. Multi-range pre-processing addresses data complexity, enhancing diagnostic accuracy and supporting clinical decision-making.