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

Q: What inputs does the two-stage multimodal framework use for OCAD diagnosis?

It uses D-SPECT images and QGS-derived functional data derived from the same imaging system.

Q: How does stage I of the framework prepare data for classification?

Stage I extracts cardiac slices along multiple axial dimensions and defines them as regions of interest (ROIs).

Q: How much did multimodal learning improve diagnostic performance compared with single-input models?

The multimodal approach achieved 81.67% accuracy, higher than D-SPECT images only (75%) and QGS-derived data only (70.00%).

Sep 30, 2026

Uploaded by นรินทร์

A two-stage multimodal learning framework is developed for automated diagnosis of obstructive coronary artery disease (OCAD) by integrating dynamic single-photon emission computed tomography (D-SPECT) images with quantitative gated SPECT (QGS)-derived functional data. Stage I extracts cardiac slices across multiple axial dimensions as regions of interest, while Stage II builds a multimodal network to extract, fuse, and classify features. A feature adaptation weighting mechanism adaptively balances modality contributions during training. Evaluation on 298 D-SPECT scans shows multimodal accuracy of 81.67%, outperforming single-input approaches, supporting a one-stop image-based clinical workflow.

LanguageEnglish

CategoryHealthcare

FormatPDF

Pages10

Reading Time25 min

AccessFree to download

Introduction

Deep learning technologies for OCAD diagnosis

Methods

Two-stage multimodal learning framework and ROI extraction

Feature adaptation weighting mechanism

Results

Diagnostic performance comparison

Conclusions

Original Article A two-stage multimodal learning framework for the automated diagnosis of obstructive coronary artery disease based on dynamic single-photon emission computed tomography Rong Wang1,2, Haijun Wang2, Chuan Zhou3, Zhongwei Li4, Dekui Chen2, Yueqian Zhu2, Jing Liu2, Xingyu Wang2, Tongtong Li5, Ping Xie1,4 1The First School of Clinical Medicine, Lanzhou University, Lanzhou, China; 2Department of Nuclear Medicine, Gansu Provincial Hospital, Lanzhou, China; 3Department of Geriatric General Surgery, Sichuan Provincial People’s Hospital, University of Electronic Science and Technology of China, Chengdu, China; 4Department of Cardiology, Gansu Provincial Hospital, Lanzhou, China; 5School of Information Science and Engineering, Lanzhou University, Lanzhou, China Contributions: (I) Conception and design: R Wang, T Li; (II) Administrative support: P Xie; (III) Provision of study materials or patients: R Wang, H Wang, C Zhou, Z Li, D Chen, Y Zhu, J Liu, X Wang; (IV) Collection and assembly of data: R Wang, H Wang, T Li; (V) Data analysis and interpretation: R Wang, T Li; (VI) Manuscript writing: All authors; (VII) Final approval of manuscript: All authors. Correspondence to: Ping Xie, MD, PhD. The First School of Clinical Medicine, Lanzhou University, No. 1 Donggang West Road, Lanzhou 730000, China; Department of Cardiology, Gansu Provincial Hospital, No. 204 Donggang West Road, Lanzhou 730000, China. Email: [1160023677@qq.com](1160023677@qq.com); Tongtong Li, CS, PhD. School of Information Science and Engineering, Lanzhou University, No. 222 South Tianshui Road, Lanzhou 730000, China. Email: [ttli2022@lzu.edu.cn](ttli2022@lzu.edu.cn). 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. Keywords: Computer-aided diagnosis; obstructive coronary artery disease (OCAD); dynamic single-photon emission computed tomography (D-SPECT); deep learning; multimodal fusion © AME Publishing Company. Quant Imaging Med Surg 2026;16(1):43 | [https://dx.doi.org/10.21037/qims-2025-617](https://dx.doi.org/10.21037/qims-2025-617)

FAQs

What inputs does the two-stage multimodal framework use for OCAD diagnosis?

It uses D-SPECT images and QGS-derived functional data derived from the same imaging system.

How does stage I of the framework prepare data for classification?

Stage I extracts cardiac slices along multiple axial dimensions and defines them as regions of interest (ROIs).

How much did multimodal learning improve diagnostic performance compared with single-input models?

The multimodal approach achieved 81.67% accuracy, higher than D-SPECT images only (75%) and QGS-derived data only (70.00%).

Sep 30, 2026

Uploaded by นรินทร์