The accurate evaluation of carotid atherosclerosis by MRI imaging is the most important factor in the assessment and management of stroke risks. Plaques can be quantified and delineated manually but this is labour-intensive and subject to error. The article describes an open-source software for the automation of carotid plaque segmentation and classification, following a hybrid approach involving the use of Med Transformer to generate high-resolution volumetric segmentation and Swin Transformer for feature-based classification. This method is more accurate, reproducible, and provides efficient carotid plaque delineation and quantification. Med Transformer attained high segmentation accuracy, with average Dice scores of 0.89 in lumen and vessel wall and 0.84 in plaque regions. Swin Transformer revealed strong performance regarding plaque type classification: the overall classification accuracy attained 91.41% and the area under the Receiver Operating Characteristic curve (AUC) was 0.9571. By fusing the results from both systems, segmentation and classification of carotid plaques could be performed under a variety of conditions and subjects with a volumetric error of less than 8%. These findings provide evidence that transformer-based systems are effective and accurate in analyzing carotid plaque in a fully automated manner, which can then be employed in scalable longitudinal studies to improve the accuracy of cerebrovascular risk assessment. The software pipeline simplifies big-data image analysis with objective and reproducible quantification. Models and scripts that are modular and developed can be integrated into clinical and research environments for further fine-tuning.

Carotid Atherosclerosis MRI Segmentation Med Transformer Swin Transformer Deep Learning Automation Medical Image Classification