Multimodal Deep Learning with Automated Myocardial Segmentation for Classification of Myocardial 99mTc-Pyrophosphate Uptake Using Planar and SPECT Imaging

Objective interpretation of myocardial 99mTc-pyrophosphate (PYP) uptake is important in suspected cardiac amyloidosis. We developed a segmentation-informed multimodal framework that classifies myocardial PYP uptake from planar and single-photon emission computed tomography (SPECT) images. The retrospective, class-enriched cohort comprised 49 studies from 44 patients (19 PYP-positive and 25 PYP-negative) with specialist-confirmed image labels. Five consecutive SPECT slices were manually selected under nuclear-medicine-physician supervision; subsequent processing was automated. SPECT Net generated an unthresholded myocardial probability map, which served as a soft, spatially interpretable representation for integration with planar features. Patient-grouped five-fold cross-validation used separate validation and test folds. In PYP-positive patients, cropped nnU-Net achieved a patient-level Dice coefficient of 0.9209±0.0316 and outperformed three segmentation comparators after Holm correction (p<0.001). Transformer-style fusion achieved 95.5% accuracy (95% confidence interval: 84.9–98.7%), 89.5% sensitivity, 100.0% specificity, and a 94.4% F1-score. Improvements over planar-only classification (90.9%; McNemar p=0.500), localized raw SPECT (86.4%; p=0.125), and H/CL assessment (88.6%; p=0.250) were not significant. These single-center internal results demonstrate the feasibility of using a soft myocardial segmentation output as an interpretable bridge for multimodal PYP uptake classification but do not establish superiority; multicenter external validation is required.