Automated Detection of Carotid Artery Encasement on CT Scans
An artificial intelligence-driven approach using fine-tuned convolutional neural networks and foundation models successfully evaluates carotid artery encasement on computed tomography scans. Developed to address the surgical challenges of head and neck squamous cell carcinomas, the method processes National Cancer Institute dataset images to measure tumor and vascular proximity.
Dataset Processing and Model Architecture
The study sourced computed tomography images and tumor masks from the National Cancer Institute’s CPTAC-HNSCC dataset. Researchers evaluated 100 patients, yielding 1,367 images that featured tumor presence. For head and neck squamous cell carcinoma segmentation, a pre-trained ResNet-50 backbone of a U-Net model was fine-tuned on the computed tomography images. Carotid segmentation utilized the SAM-Med2D foundation model. Both models returned segmentations for the tumor and the carotid, respectively.
Validation Scores and Measurement Metrics
Encasement was assessed by calculating the minimum distance and encasement angle between the masks provided by the models. During validation, the tumor segmentation model achieved a mean F1 score of 86 percent and an Intersection over Union score of 75 percent. Specifically, it reached a 73 percent Intersection over Union score for the tumor class and 99 percent for the non-tumor class. Carotid segmentation proved most effective when using the bounding-box method, reaching 79 percent accuracy. Both models were successfully applied to the same image for the final assessment.
Radiomics and Deep Learning Comparisons
In a related evaluation of carotid artery disease on computed tomography angiography images published by Elsevier, researchers compared radiomics and deep learning methods against the conventional calcium score. Analyzing 132 carotid arteries—split into 41 culprit, 41 non-culprit, and 50 asymptomatic arteries—the study found that radiomics attained a mean area under the curve of 0.96 for asymptomatic versus symptomatic arteries. Deep learning reached 0.86, while the calcium score achieved 0.79. For culprit versus non-culprit arteries, radiomics achieved a mean area under the curve of 0.75, followed by deep learning at 0.67 and the calcium score at 0.60. Multi-class classification mean area under the curve scores were 0.95 for radiomics, 0.79 for deep learning, and 0.71 for the calcium score. Explainability algorithms including SHAP and GRAD-CAM revealed consistent patterns in the most important radiomic features.
