An Automated Deep Learning Framework for Thyroid Contrast-Enhanced Ultrasound Video Malignancy Assessment

Abstract

Contrast-enhanced ultrasound (CEUS) is vital for diagnosing thyroid nodule malignancy, but current time-intensity curve (TIC) analysis is time-consuming and subjective due to manual steps. This study introduces SPAF (Thyroid CEUS Segmentation and Peak Intensity Frame Analysis Framework), a novel deep learning framework designed to automate the entire CEUS evaluation process: nodule segmentation, TIC analysis, and malignancy classification. SPAF integrates an enhanced YOLOv5 segmentation network and a novel peak intensity frame (PIF) extraction algorithm to overcome traditional inefficiencies. Validated on a multicenter cohort of 600 patients, SPAF demonstrated a 4 to 14-fold increase in PIF extraction efficiency over commercial software (QLAB). Crucially, the PIF-based malignancy prediction achieved by SPAF was comparable to or superior to QLAB, demonstrating strong performance with AUCs up to 0.82 on valid set, respectively. This automated and accurate framework provides a powerful alternative to manual quantitative CEUS analysis, significantly reducing sonographers' workload and enhancing diagnostic consistency.

Ähnliche Arbeiten