Detection of myocardial ischemia using convolutional neural network (CNN) on magnetic resonance perfusion images

Purpose: Stress cardiac magnetic resonance imaging (CMRI), typically using adenosine, is a leading non-invasive tool to detect significant coronary artery disease (CAD), with around 90% sensitivity. This study aimed to evaluate the feasibility of a convolutional neural network (CNN)-based approach to detect coronary stenoses requiring intervention through adenosine stress perfusion CMRI, validated against coronary angiography.

Methods: This retrospective study analyzed 300 perfusion image slices (apical, medial, basal) from 100 patients (mean age 67 ± 11 years; 82% male) who underwent adenosine stress CMRI on a 1.5T MRI scanner and coronary angiography within 90 days between 2016 and 2024. Using the 16-segment AHA model (Figure), ischemia was validated through coronary angiography, with 63 ischemic and 37 non-ischemic patients. The dataset was split into training (70%), validation (15%), and testing (15%) sets, with 40% positive labels in each subset. CMRI images were pre-processed by cropping to 128×128 pixels centered on the heart. Time-resolved image sequences were processed to extract 10 key frames starting from contrast flooding. Myocardium localization in the first frame was achieved using the YOLOv4 object detection network trained on annotated CMRI images. Intensity changes within a region of interest (ROI) surrounding the myocardium were detected to identify the frame of contrast flooding. The 10 frames were used as 3D patches for training CNNs. Two training approaches were tested: (1) composing the 10 frames into 128×128×10 for 2D CNNs, and (2) depth-wise composing them into a 3D input (10×128×128×3) for 3D CNNs. Performance of SE-Net154 (with Squeeze-and-Excitation blocks) and ResNet101 was compared.

Results: Table summarizes comparative measures of the diagnostic performance of the four investigated classifiers determined based on receiver operating characteristic curve (ROC). Overall, the 3D SE-Net154 stands out as the most balanced model in terms of performance with a relatively high specificity (80.77%) and a moderate sensitivity (47.37%), reflected in an overall AUC of 63%, which is the highest among the classifiers. The 2D ResNet101 shows a more balanced trade-off between sensitivity (57.89%) and specificity (61.54%), but its AUC of 58% suggests room for improvement. The 2D SE-Net154 has the highest sensitivity (94.74%), making it highly effective at identifying true positives, but its specificity (15.38%) is extremely low. With an AUC of 47%, it doesn't perform well overall due to its high Type I error (84.62%), despite its strength in sensitivity. 3D ResNet101 has the weakest performance based on AUC, indicating that it struggles to balance sensitivity and specificity effectively.

Conclusions: This study presents an AI-based approach for detecting coronary stenoses via adenosine stress perfusion CMRI, potentially reducing unnecessary invasive procedures. While none of the classifiers showed strong performance, the 3D SE-Net154 offered the best balance between sensitivity and specificity. Future work will focus on incorporating attention mechanisms in CNNs to enhance diagnostic accuracy and improve the detection of ischemia on CMRI.