Carotid Plaque Segmentation in Ultrasound Images Using a Mask R-CNN

Ultrasound imaging plays a pivotal role in diagnosing carotid atherosclerosis, a significant precursor to cardiovascular and cerebrovascular diseases and events. This noninvasive modality provides real-time, high-resolution images, allowing clinicians to assess atherosclerotic plaques in the carotid arteries without invasive procedures. Early detection using ultrasound aids in timely interventions, reducing the risk of adverse cardiovascular events. Purpose: In this study, we present the refinement of a Mask R-CNN model initially designed for carotid lumen detection to automatically generate bounding boxes (BB) enclosing atherosclerotic plaque for segmentation to assist in our ultrasound elastography workflow.

We utilize a PyTorch torchvision implementation of the Mask R-CNN for carotid plaque detection and BB placement. Our dataset consists of 118 severe stenotic carotid plaques from presenting patients, clinically indicated for a carotid endarterectomy. Due to the variability of plaque presentation in the dataset, a multitude of different R-CNN models were observed to have varying results based on the allowed number of prediction regions. An overview analysis looking at shared predictions from these models showed a slight improvement compared to the individual model results.

Evaluation metrics such as Dice similarity coefficient and intersection over Union are employed. The model trained with 5 maximum BB prediction regions and tested with 2 maximum BB prediction regions produced the highest individual accuracy with a Dice score of 0.74 and intersection over union of 0.61. A filtered combined analysis of all the models demonstrated a slight increase in performance with scores of 0.76 and 0.61 respectively.

Due to the significant variation in plaque presentation and types amongst presenting patients, the accuracy of the Plaque Mask R-CNN network would benefit from the incorporation of additional patient datasets to incorporate increased variation into the training dataset.