Prediction of Central Lymph Node Metastasis in Papillary Thyroid Microcarcinoma Using a Deep Learning Radiomics Model...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Papillary thyroid microcarcinoma (PTMC) diagnosis and staging are crucial.
• Accurate segmentation of PTMC on ultrasound images is challenging.
• Predicting central lymph node metastasis (CLNM) noninvasively is clinically significant.

Purpose of the Study:

• To establish a Segment Anything Model 3 (SAM3) for automatic PTMC segmentation using ultrasound.
• To develop and validate a deep learning radiomics (DLR) model for noninvasive CLNM prediction in PTMC.
• To assess the clinical utility of these AI models in PTMC management.

Main Methods:

• Retrospective collection of ultrasound data from 1999 PTMC patients across multiple centers.
• Automatic tumor segmentation using SAM3 on PTMC ultrasound images.
• Extraction of deep learning and radiomics features, followed by DLR model development and validation.

Main Results:

• SAM3 achieved high segmentation performance with Dice similarity coefficients of 0.882 in validation and 0.859 in external testing.
• The DLR model demonstrated superior performance in predicting CLNM, with AUCs of 0.901 (training), 0.875 (validation), and 0.853 (external testing).
• Decision curve analysis confirmed the clinical utility of the DLR model.

Conclusions:

• SAM3-based DL segmentation provides satisfactory PTMC delineation on ultrasound.
• The DLR model offers a noninvasive approach for predicting CLNM in PTMC.
• These AI tools can support clinical decision-making for PTMC patients.