Classification of Thyroid Nodules by Using Deep Learning Radiomics Based on Ultrasound Dynamic Video

Abstract

OBJECTIVES: We aimed to design a radiomics model for differential diagnosis of thyroid carcinoma based on dynamic ultrasound video, and compare its diagnostic performance with that of radiomics model based on static ultrasound images. METHODS: Between January 2019 and May 2021, 890 patients with 1015 thyroid nodules (775 for training, 240 for validation) were prospectively enrolled. In total 890 patients underwent thyroidectomy within 1 month, and ultrasound dynamic video and static images were both acquired. Two deep learning (DL) models, namely DL-video and DL-image models, were proposed to diagnose thyroid nodules by analyzing ultrasound video and static images respectively. The performance of models was assessed by areas under the receiver operating characteristic curve (AUC). The DL model on ultrasound cines was re-visualized to help radiologists understand its potential working mechanism. RESULTS: The AUC of DL-video were 0.947 (95% CI: 0.931-0.963) and 0.923 (95% CI: 0.892-0.955) in training and validation cohorts, respectively. For DL-image model, the AUC were 0.928 (95% CI: 0.910-0.945) and 0.864 (95% CI: 0.819-0.910), respectively. The diagnosis performance of the DL-video was superior to that of DL-image, and there was significant difference between the AUC of DL-video and DL-image model in validation cohort (P = .028). The visualization demonstrated certain important ultrasound features that could be recognized by human eyes. CONCLUSIONS: The proposed DL radiomics model based on dynamic ultrasound video can accurately and individually classified thyroid nodules. The constructed DL-video model combining ultrasound video holds good potential for benefiting the management of patients with thyroid nodules.

Ähnliche Arbeiten