A deep learning-based automated diagnosis system for SPECT myocardial perfusion imaging

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Myocardial perfusion imaging single-photon emission computed tomography (MPI SPECT) is crucial for diagnosing cardiovascular diseases.
• Image noise and artifacts in MPI SPECT can hinder accurate diagnosis.
• Deep learning offers potential solutions for improving image analysis and diagnostic accuracy.

Purpose of the Study:

• To develop and evaluate a deep learning-based diagnosis support system for MPI SPECT images.
• To assess the system's performance in classifying coronary artery disease (CAD) and visualizing lesions.
• To compare the diagnostic accuracy and efficiency of AI-guided diagnosis versus traditional cardiologist diagnosis.

Main Methods:

• A dataset of 5443 MPI SPECT images was collected and labeled for healthy or CAD.
• A convolutional neural network (CNN) was trained on reconstructed 3D/4D MPI SPECT data.
• Unsupervised learning and Grad-CAM were used for lesion visualization and AI model interpretability.

Main Results:

• The trained CNN achieved high performance (AUC ROC: ~0.91, AUC R-P: 0.87).
• The AI-based automated diagnosis system achieved 88% accuracy, outperforming cardiologists alone (65%) and AI-guided cardiologists (80%).
• AI-guided diagnosis significantly reduced diagnosis time (12 min) compared to cardiologists alone (31 min).

Conclusions:

• A deep learning-based diagnosis support system can significantly improve diagnostic accuracy for MPI SPECT.
• The AI system effectively visualizes diseased lesions, aiding cardiologists in diagnosis.
• This AI tool has the potential to enhance cardiologist efficiency and patient care in cardiovascular disease diagnosis.