Segmentation-guided peritumoral radiomics and machine learning for ultrasound-based classification of salivary gland tumors.
Authors
Affiliations (7)
Affiliations (7)
- Department of Otolaryngology, Head and Neck Surgery, Far Eastern Memorial Hospital, New Taipei City, Taiwan.
- Department of Endocrinology, Far Eastern Memorial Hospital, New Taipei City, Taiwan.
- Department of Communication Engineering, Asia Eastern University of Science and Technology, New Taipei City, Taiwan.
- Head and Neck Cancer Surveillance and Research Study Group, Far Eastern Memorial Hospital, New Taipei City, Taiwan.
- Department of Biomedical Engineering, National Yang Ming Chiao Tung University, Taipei, Taiwan.
- Graduate Institute of Medicine, Yuan Ze University, Taoyuan, Taiwan.
- Department of Electrical Engineering, Yuan Ze University, Taoyuan, Taiwan.
Abstract
Preoperative differentiation of benign and malignant salivary gland tumors (SGTs) remains challenging because of overlapping ultrasonographic features. To develop and validate a segmentation-guided machine learning model using peritumoral radiomic features for ultrasound-based classification of SGTs. A YOLOv8 segmentation model was developed using 366 ultrasound images collected between January 2007 and June 2022. Automated tumor segmentation generated a standardized peritumoral ribbon for extracting 15 radiomic features characterizing tumor morphology, boundary microtexture, and boundary-to-core properties. After correlation-based feature selection, five machine learning classifiers were developed and evaluated in an independent testing cohort (<i>n</i> = 121; July 2022-December 2024) and an internal validation cohort (<i>n</i> = 51; January-December 2025). The YOLOv8 model achieved Dice coefficients of 0.9485 and 0.8896 in the training and testing cohorts, respectively. Twelve radiomic features were retained for model development. Among the evaluated classifiers, the Support Vector Classifier (SVC) demonstrated the best performance, achieving AUCs of 0.8117 and 0.7951 in the testing and validation cohorts, respectively. The complete pipeline was implemented as both standalone and web-based applications. Segmentation-guided peritumoral radiomics combined with machine learning represents an interpretable and promising adjunctive approach for ultrasound-based classification of salivary gland tumors.