Malignancy prediction for calcified thyroid nodules using deep learning based on ultrasound dynamic videos

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Calcification, particularly microcalcification, is a key indicator of malignancy risk and is linked to papillary thyroid carcinoma (PTC).
• Current ultrasound (US) imaging struggles to detect subtle calcifications, leading to delayed PTC treatment or unnecessary procedures.

Purpose of the Study:

• Develop a deep learning (DL) model optimized for calcification detection using dynamic ultrasound videos.
• Enhance the accuracy and interpretability of diagnosing calcified thyroid nodules to determine malignancy.

Main Methods:

• Retrospective collection of 2,319 dynamic ultrasound videos from 1,257 patients across six hospitals.
• Development of a 3D InceptionResNetV2 DL model with a calcification attention module for enhanced micro-calcification sensitivity.
• Comparison of model performance against static 2D ultrasound images and diagnoses from junior and senior radiologists.

Main Results:

• The optimized 3D DL model achieved an AUROC of 0.916, sensitivity of 0.860, and specificity of 0.834 on the external test set.
• The model significantly outperformed radiologists (AUROC 0.916 vs. 0.638) and improved their diagnostic accuracy.
• Interpretable heatmaps from 3D Grad-CAM confirmed the model's focus on calcified regions, aligning with clinical logic.

Conclusions:

• A calcification-optimized DL model trained on dynamic ultrasound videos can efficiently predict the malignancy of calcified thyroid nodules.
• This interpretable, non-invasive tool shows potential for early PTC detection, improving diagnosis and treatment planning.