Deep learning-based early prediction of carotid plaque response to lipid-lowering therapy using longitudinal multimodal ultrasound imaging.
Authors
Affiliations (5)
Affiliations (5)
- Department of Ultrasound Imaging, The First People's Hospital of Wenling (Taizhou University Affiliated Wenling Hospital), School of Medicine, Taizhou University, Taizhou, People's Republic of China.
- Department of Special Examination, Haining Central Hospital, Haining City, People's Republic of China.
- Wenling Institute of Big Data and Artificial Intelligence in Medicine, Taizhou, People's Republic of China. [email protected].
- Faculty of Applied Sciences, Macao Polytechnic University, Macao, People's Republic of China. [email protected].
- Department of Ultrasound Imaging, The First People's Hospital of Wenling (Taizhou University Affiliated Wenling Hospital), School of Medicine, Taizhou University, Taizhou, People's Republic of China. [email protected].
Abstract
This study aimed to develop and validate a deep learning prediction model using longitudinal multimodal ultrasound imaging for early identification of treatment-sensitive and treatment-resistant carotid plaques in patients receiving lipid-lowering therapy. This prospective study enrolled 802 patients with vulnerable carotid plaques or stenosis ≥ 50%. Patients underwent serial multimodal ultrasound examinations, including B-mode imaging, superb microvascular imaging, and shear wave elastography at baseline and 3, 6, 9, and 12 months after initiating statin therapy. The dataset was divided into training and testing sets using stratified sampling with data augmentation. A hybrid DL model combining convolutional neural networks and long short-term memory networks analyzed longitudinal imaging sequences integrated with baseline clinical data. Five progressive prediction models were constructed for baseline and each follow-up time point, sharing identical architecture but trained independently on temporal sequences of varying lengths using 5-fold cross-validation. Model performance was assessed for discrimination ability, calibration consistency, and clinical utility. Five progressive prediction models demonstrated characteristic temporal performance patterns, with significant improvement from 3 to 6 months (AUC 0.866), followed by marginal gains. The 6-month model emerged as the most clinically practical assessment time point, achieving high specificity (93.7%) for early therapeutic decisions. Ablation experiments confirmed imaging features as primary predictive determinants, while attention mapping revealed consistent focus on plaque-adjacent regions, validating that treatment response prediction relies on morphological changes within target plaques. A hybrid DL model enables reliable carotid plaque treatment response prediction within six months, optimizing personalized therapy through earlier identification of treatment-resistant patients. This study validates deep learning algorithms to predict carotid plaque treatment response within six months, advancing clinical radiology practice by enabling earlier therapeutic optimization through objective ultrasound-based assessment. Conventional imaging requires 12 months to reliably assess plaque treatment response. Deep learning model predicts treatment response at six months with high accuracy. Earlier prediction enables timely therapeutic adjustments for resistant patients.