Carotid Plaque Segmentation and Classification using MRI-based Plaque Texture Analysis and Convolutional Neural Network

Zakrya Hasan Abu Alregal¹Gehad Abdullah Amran^2*Ali A. Al-Bakhrani²SALEH ABDUL AMIR MOHAMMAD²Amerah Alabrah³Lubna Alkhalil³Abdalla Ibrahim⁴

• ¹Central South University, Changsha, 410083, China, Changsha, China
• ²Dalian University of Technology, Dalian, China
• ³King Saud University, Riyadh, Riyadh, Saudi Arabia
• ⁴Central South University, Changsha, Hunan Province, China

Background: Accurate segmentation and classification of carotid plaques are critical for assessing stroke risk. However, conventional methods are hindered by manual intervention, inter-observer variability, and poor generalizability across heterogeneous datasets, limiting their clinical utility. Methods: We propose a hybrid deep learning framework integrating Mask R-CNN for automated plaque segmentation with a dual-path classification pipeline. A dataset of 610 expert-annotated MRI scans from Xiangya Hospital was processed using Plaque Texture Analysis Software (PTAS) for ground truth labels. Mask R-CNN was fine-tuned with multi-task loss to address class imbalance, while a custom 13-layer CNN and Inception V3 were employed for classification, leveraging handcrafted texture features and deep hierarchical patterns. The custom CNN was 1 Sample et al. Running Title evaluated via K10 cross-validation, and model performance was quantified using Dice Similarity Coefficient (DSC), Intersection over Union (IoU), accuracy, and ROC-AUC. Results: The Mask R-CNN achieved a mean DSC/IoU of 0.34, demonstrating robust segmentation despite anatomical complexity. The custom CNN attained 86.17% classification accuracy and an ROC-AUC of 0.86 (p = 0.0001), outperforming Inception V3 (84.21% accuracy).Both models significantly surpassed conventional methods in plaque characterization, with the custom CNN showing superior discriminative power for high-risk plaques.This study establishes a fully automated, hybrid framework that synergizes segmentation and classification to advance stroke risk stratification. By reducing manual dependency and inter-observer variability, our approach enhances reproducibility and generalizability across diverse clinical datasets. The statistically significant ROC-AUC and high accuracy underscore its potential as an AI-driven diagnostic tool, paving the way for standardized, data-driven cerebrovascular disease management.