Cross‑modal image‑text collaborative deep‑learning model for predicting severe coronary stenosis from CCTA: a retrospective single‑center study

Background Coronary computed tomography angiography (CCTA) is a cornerstone non‑invasive modality for coronary artery evaluation. Nevertheless, conventional manual reading suffers from subjective bias and limited reproducibility. We proposed a cross‑modal deep‑learning framework integrating image and clinical text information to improve detection of severe coronary stenosis. Objective To develop and validate a cross‑modal radiomics deep‑learning model based on CCTA data and further evaluate its clinical performance for identifying severe coronary artery stenosis. Methods In this single‑center retrospective study, 96 patients receiving both CCTA and invasive coronary angiography (ICA) were enrolled, yielding 288 raw coronary vessel‑level samples; two vessels were excluded for vessel‑specific poor image quality, resulting in 286 evaluable coronary vessel‑level (unit of analysis) samples. A dual‑branch cross‑modal model combining ResNet‑18 image encoder and lightweight Transformer text encoder was constructed. Class‑differentiated augmentation and multi‑loss joint optimization were applied to mitigate class imbalance. Dataset was split into training/test set (8:2). Using ICA as reference gold‑standard, ROC‑based diagnostic metrics (AUC, sensitivity, specificity, accuracy, PPV, NPV) were calculated for model and manual interpretation. Results Our cross‑modal deep‑learning model achieved AUC = 0.83 (95%CI 0.71‑0.94), sensitivity 64.0%, specificity 93.0%, accuracy 79.0%, PPV 89.0%, NPV 74.0%. It outperformed manual CCTA reading (AUC = 0.676; sensitivity 46.38%, specificity 88.7%, accuracy 68.5%). Conclusion The proposed cross‑modal image‑text collaborative deep‑learning model improves diagnostic performance for severe coronary stenosis. It could serve as a promising intelligent auxiliary tool for CCTA‑based non‑invasive assessment of coronary artery disease.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-18
DOI
https://doi.org/10.21203/rs.3.rs-11431659/v1
Primary Topic
Cardiac Imaging and Diagnostics
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Cross‑modal image‑text collaborative deep‑learning model for predicting severe coronary stenosis from CCTA: a retrospective single‑center study

Haibao Yuan, Li Ma, Huan Tang, Dongdong Tang et al.
4 links
Research Square
Cardiac Imaging and Diagnostics
preprint

Cross‑modal image‑text collaborative deep‑learning model for predicting severe coronary stenosis from CCTA: a retrospective single‑center study

Haibao Yuan, Li Ma, Huan Tang, Dongdong Tang, Linghui Zhao, Jiangpeng Wang
preprint en

Abstract

Background Coronary computed tomography angiography (CCTA) is a cornerstone non‑invasive modality for coronary artery evaluation. Nevertheless, conventional manual reading suffers from subjective bias and limited reproducibility. We proposed a cross‑modal deep‑learning framework integrating image and clinical text information to improve detection of severe coronary stenosis. Objective To develop and validate a cross‑modal radiomics deep‑learning model based on CCTA data and further evaluate its clinical performance for identifying severe coronary artery stenosis. Methods In this single‑center retrospective study, 96 patients receiving both CCTA and invasive coronary angiography (ICA) were enrolled, yielding 288 raw coronary vessel‑level samples; two vessels were excluded for vessel‑specific poor image quality, resulting in 286 evaluable coronary vessel‑level (unit of analysis) samples. A dual‑branch cross‑modal model combining ResNet‑18 image encoder and lightweight Transformer text encoder was constructed. Class‑differentiated augmentation and multi‑loss joint optimization were applied to mitigate class imbalance. Dataset was split into training/test set (8:2). Using ICA as reference gold‑standard, ROC‑based diagnostic metrics (AUC, sensitivity, specificity, accuracy, PPV, NPV) were calculated for model and manual interpretation. Results Our cross‑modal deep‑learning model achieved AUC = 0.83 (95%CI 0.71‑0.94), sensitivity 64.0%, specificity 93.0%, accuracy 79.0%, PPV 89.0%, NPV 74.0%. It outperformed manual CCTA reading (AUC = 0.676; sensitivity 46.38%, specificity 88.7%, accuracy 68.5%). Conclusion The proposed cross‑modal image‑text collaborative deep‑learning model improves diagnostic performance for severe coronary stenosis. It could serve as a promising intelligent auxiliary tool for CCTA‑based non‑invasive assessment of coronary artery disease.

Research Square
Xuzhou Medical College (CN), The Fifth People's Hospital of Anyang (CN), Second People’s Hospital of Huai’an (CN)
Good health and well-being
Cardiac Imaging and Diagnostics
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon