Reliable deep learning for coronary artery disease detection: a patient-level, statistically validated MRI study

BACKGROUND: Accurate detection of coronary artery disease (CAD) from cardiac magnetic resonance (CMR) imaging can support earlier diagnosis and streamlined clinical decision-making. OBJECTIVE: This study evaluated the performance and statistical robustness of two deep learning architectures - DenseNet121 and ResNet50 - for automated CAD classification using multiparametric CMR imaging. METHODS: Images were preprocessed using a valid pipeline and partitioned strictly at the patient level. Model performance was quantified through average accuracy, area under the receiver operating characteristic (ROC) curve (AUC-ROC), precision recall, while distributional assumptions were assessed using Shapiro-Wilk tests and variance homogeneity was explored with Brown-Forsythe test. RESULTS: ResNet50 demonstrated the strongest performance, achieving an average accuracy of 90.43%, AUC-ROC of 0.862, and area under the precision recall curve (PR-AUC) of 0.891. DenseNet121 showed lower accuracy (81.72%). Statistical analysis revealed non-normal performance distributions and significant variance differences between models. CONCLUSIONS: The findings indicate that ResNet50 offers a reliable and statistically validated solution for CAD detection from CMR imaging. The combined use of realistic preprocessing and comprehensive inferential testing supports the generation of reproducible and clinically meaningful performance estimates.