A two-stage multimodal learning framework for the automated diagnosis of obstructive coronary artery disease based on dynamic single-photon emission computed tomography.

Background: Obstructive coronary artery disease (OCAD) is among the most life-threatening cardiovascular diseases in the world. Dynamic single-photon emission computed tomography (D-SPECT) offers a noninvasive technique for visualizing the perfusion of the heart and the functional state of the myocardium. However, computer-aided diagnostic models for OCAD mainly focus on analyzing medical images and neglect the potential benefits of integrating functional parameters derived from quantitative gated single-photon emission computed tomography (QGS) obtained within the same imaging system. The objective of this study was to develop a two-stage multimodal learning framework that integrates D-SPECT images and QGS-derived functional data to enhance diagnostic accuracy and support a one-stop, imaging-based diagnostic workflow for OCAD. Methods: We developed a two-stage multimodal learning framework for automated OCAD diagnosis using both D-SPECT images and QGS-derived functional data. In stage I, cardiac slices along multiple axial dimensions were extracted as regions of interest (ROIs). In stage II, a multimodal learning network was constructed to extract, fuse, and classify features from both inputs. Furthermore, a feature adaptation weighting mechanism (FAWM) was introduced to adaptively allocate the contributions of different modalities during training. Finally, the performance of the proposed model was evaluated on a dataset of 298 D-SPECT scans. Results: The multimodal method achieved an accuracy of 81.67%, outperforming models trained with single inputs (D-SPECT images only: 75%, P=0.12; QGS-derived data only: 70.00%, P<0.05; paired t-test). Conclusions: The integration of D-SPECT imaging and QGS-derived functional parameters within a unified multimodal framework significantly improves diagnostic performance for OCAD. This approach demonstrates the feasibility of serving as a one-stop, image-based diagnostic workflow for clinical practice.