Robust Amyloidosis Subtype Classification via Multisequence CMR Fusion with Spatiotemporal Learning

Parker Martin, Yifan Liu, Karolina Zareba, Suzanne Smart, Akash Goyal, Orlando Simonetti, Jeremy Slivnick, Yuan Xue

Keywords: Cardiac Amyloidosis, Cardiac MRI, Deep Learning, Multimodal Learning, Spatiotemporal Modeling, Model Interpretability

TL;DR: We present a multimodal model integrating cine, LGE, and mapping sequences from cardiac MRI using spatiotemporal learning and gated attention for robust, interpretable amyloidosis subtype classification allowing incomplete input sequences.

Abstract: Cardiac amyloidosis (CA) subtype classification remains a critical diagnostic challenge. We propose a multimodal deep learning framework that integrates cine, late gadolinium enhancement (LGE), and T1/T2 parametric cardiac MRI sequences to differentiate light chain (AL) and transthyretin (ATTR) amyloidosis. The model employs sequence-specific encoders and gated attention fusion, enabling robust performance even with missing input sequences. Evaluated on 123 patients with cross-validation, the xLSTM-based model achieved the highest AUC (0.8506), outperforming a Video Swin Transformer (VST) alternative. Grad-CAM visualizations highlight both cardiac and extracardiac regions, demonstrating interpretability and the potential for identifying systemic imaging biomarkers. These results support a clinically viable approach to non-invasive CA subtype diagnosis.

Submission Number: 99