Comparative Evaluation of U-Net Variants for Automated Segmentation of Cardiac SPECT Images Using Diverse Encoder Architectures

Introduction Myocardial perfusion imaging (MPI) using single-photon emission computed tomography (SPECT) is widely used for diagnosing and risk-stratifying coronary artery disease (CAD). Accurate myocardial segmentation is essential for reliable quantification of left ventricular function; however, manual delineation remains time-consuming and prone to inter-observer variability. This study systematically compares five U-Net encoder backbones (CNN, MobileNet, VGG-Attention, Inception, and ResNet50-Attention) and proposes a novel training-free weighted ensemble strategy suitable for real-time clinical deployment. Methods Two MPI SPECT datasets were analyzed (small: 609 images; large: 5,169 images), each paired with expert-verified ground truth masks. Images were preprocessed to 64×64 pixels and augmented using flipping, grid distortion, Gaussian noise injection, and contrast adjustments. All models were trained under identical hyperparameter settings and evaluated using mIoU, Dice coefficient, pixel accuracy, and AUC. K-fold cross-validation (K=5) and an ablation study comparing cross-entropy and Tversky loss functions were performed to assess statistical robustness and segmentation balance. Results The Inception-based U-Net achieved the highest single-model performance on both datasets (small: Dice 0.9656, mIoU 0.9358; large: Dice 0.9538, mIoU 0.9197). The proposed weighted ensemble consistently outperformed all individual architectures (small: Dice 0.9679, mIoU 0.9398, AUC 0.9994, pixel accuracy 99.45%; large: Dice 0.9544, mIoU 0.9205, AUC 0.9979, pixel accuracy 99.25%), achieving near-zero false-positive rates and the lowest false-negative rates across both datasets. K-fold cross-validation confirmed improved foreground detection robustness over holdout splitting. MobileNet offers the best computational efficiency (training time: 3.71 min; model size: 24.04 MB), representing a viable option for resource-constrained settings. Conclusion The proposed training-free weighted ensemble of U-Net architecture achieves state-of-the-art segmentation accuracy for myocardial perfusion SPECT imaging, with near-perfect AUC and pixel accuracy across datasets of varying scale. These results demonstrate strong potential for clinical integration into nuclear cardiology workflows, reducing dependence on manual segmentation and supporting efficient, reproducible left ventricular quantification.