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

Background/Objectives: Single-photon emission computed tomography (SPECT) myocardial perfusion imaging (MPI) is widely used for non-invasive assessment of coronary artery disease under stress and rest conditions. Although deep learning has shown promise for automated SPECT MPI interpretation, most studies focus on single-tracer datasets and do not explicitly account for tracer-dependent variability. This study developed and evaluated a multi-task deep learning framework with tracer-specific prediction heads for patient-level SPECT MPI classification. Methods: A convolutional neural network with a shared feature encoder and tracer-specific heads was implemented using polar map representations from technetium-99m (Tc-99m) and thallium-201 (Tl-201) studies. Transfer learning from ImageNet was applied. Stress-only, rest-only, and dual-input configurations were evaluated using repeated patient-stratified cross-validation and independent testing. Performance was assessed using ROC-AUC and balanced accuracy. Results: For Tc-99m normal versus abnormal perfusion classification, the stress-only model achieved the highest cross-validation AUC (0.88 ± 0.067) and test AUC of 0.88 [0.67–0.99]. For Tl-201 low-risk versus intermediate/high-risk classification, stress-based models achieved the highest cross-validation AUC (0.88 ± 0.051) and test AUC of 0.80 [0.71–0.89], comparable to dual-input models. In both tracer-specific tasks, stress-phase information showed favorable performance, but the endpoints differed and should be interpreted separately. Conclusions: Stress-phase polar maps provided strong discriminative information within this single-center cohort. These findings should be interpreted in a tracer- and task-specific manner supporting stress-phase imaging as an informative input for AI-based SPECT MPI classification while underscoring the need for external validation before broader clinical generalization.