Atherosclerosis diagnosis from magnetic resonance images: an empirical approach to automate carotid vessel wall segmentation and feature-based prediction.
Authors
Affiliations (7)
Affiliations (7)
- Department of Electrical and Electronic Engineering, University of Dhaka, Dhaka, 1000, Bangladesh.
- Department of Electrical Engineering, Qatar University, Doha, 2713, Qatar.
- Department of Radiology, Hamad Medical Corporation, Doha, Qatar.
- Intelligent Signal Processing (ISP) Research Lab, Department of Electronics and Communication Engineering, Kuwait College of Science and Technology, Block 4, Doha, Kuwait. [email protected].
- Department of Electronics and Communication Engineering, Vels Institute of Sciences, Technology, and Advanced Studies, Chennai, Tamilnadu, India. [email protected].
- Faculty of Electronic Engineering and Technology, Universiti Malaysia Perlis, Arau, Perlis, Malaysia.
- Department of Electrical Engineering, Qatar University, Doha, 2713, Qatar. [email protected].
Abstract
Carotid artery atherosclerosis is a severe vascular condition linked to stroke and cardiovascular complications, requiring early and precise diagnosis. Manual segmentation of carotid artery structures in MRI is time-consuming and prone to interobserver variability. This study proposes and validates an automated framework combining deep learning (DL)-based segmentation with machine learning (ML) classification for diagnosing carotid atherosclerosis from low-resolution MRI VISTA images. We used 30 patient cases from the COSMOS 2022 MICCAI dataset, comprising 3238 annotated 2D axial slices from 3D scans. A subject-independent five-fold cross-validation strategy was employed, with 647 test images per fold. Data imbalance was addressed using augmentation and class balancing. All scans were acquired using a 3T Philips MRI system with a 3D VISTA black-blood sequence (TR = 800 ms, TE = 20 ms, resolution = 0.6 × 0.6 × 0.6 mm³). Expert-guided annotations were performed by trained researchers. A DenseNet121-UNet++-based model, enhanced through ablation-driven improvements, was used for segmentation, evaluated using Dice Similarity Coefficient (DSC) and Intersection over Union (IoU). The proposed method achieved a DSC of 85.76% and IoU of 73.68%. For classification, 466 handcrafted features were extracted and ranked using XGBoost, Random Forest, and Extra Trees, followed by stacking ensemble learning. The classification accuracy improved to 86.94%, outperforming individual models. The results demonstrate that improved segmentation significantly enhances classification performance. The proposed hybrid DL-ML pipeline provides an accurate, interpretable, and clinically feasible solution for automated carotid atherosclerosis diagnosis.