Decoding the Myocardium: Tracer-Aware Deep Learning for Patient-Level Classification in Stress-Rest SPECT Myocardial...

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Single-photon emission computed tomography (SPECT) myocardial perfusion imaging (MPI) is crucial for assessing coronary artery disease.
• Existing deep learning models for SPECT MPI often overlook tracer-specific variations.
• Automated interpretation of SPECT MPI requires methods that account for different radiotracers.

Purpose of the Study:

• To develop and assess a multi-task deep learning framework for patient-level SPECT MPI classification.
• To incorporate tracer-specific prediction heads to handle variability between technetium-99m (Tc-99m) and thallium-201 (Tl-201).
• To evaluate the performance of stress-only, rest-only, and dual-input models.

Main Methods:

• A convolutional neural network with a shared encoder and tracer-specific heads was used.
• Polar map representations of SPECT MPI data were utilized.
• Transfer learning from ImageNet and patient-stratified cross-validation were employed.

Main Results:

• Stress-only models demonstrated high performance for both Tc-99m (test AUC 0.88) and Tl-201 (test AUC 0.80) classification tasks.
• Stress-phase information proved highly discriminative for SPECT MPI interpretation.
• Performance varied based on tracer and specific classification endpoint.

Conclusions:

• Stress-phase SPECT MPI polar maps contain significant discriminative information for AI-based classification.
• The developed framework shows promise for automated SPECT MPI interpretation, accounting for tracer differences.
• Further external validation is necessary for broad clinical generalization of these AI models.