Automated identification of cardiac amyloidosis using cross-modal neural networks on [Formula: see...

Accurate, early identification of transthyretin cardiac amyloidosis (ATTR-CA) is challenging yet critical for effective treatment. In this work, the modeled endpoint is scan positivity on [Formula: see text] scintigraphy, defined by the semi-quantitative Perugini visual grade (Grade 2-3 versus Grade 0-1). Two multimodal deep-learning frameworks, including Late Fusion (LF) and Cross-Modal Fusion Network (CMF-Net), are proposed to combine [Formula: see text] scintigraphy with clinical metadata for automated detection. On a curated cohort of 109 patients (62 positive, 47 negative), fusion models consistently outperformed image-only convolutional neural networks (CNNs): CMF-Net raised average accuracy by 6.9 percentage points and LF by 5.4. EfficientNet CMF-Net achieved peak accuracy 90.9% and F1-score 91.4%. Notable gains included a +30.8 percentage points sensitivity(recall) for ResNet-34 with CMF-Net and ResNet-50 LF sensitivity of 94.9%. These results show that integrating imaging and text/numeric clinical data yields superior, reproducible detection of [Formula: see text] scan positivity and may streamline scan interpretation.