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

## TL;DR

A deep learning model using dynamic ultrasound videos improves the accuracy of predicting malignancy in calcified thyroid nodules compared to radiologists and static imaging.

## Contribution

A calcification-optimized 3D deep learning model using dynamic ultrasound videos for thyroid nodule malignancy prediction is proposed.

## Key findings

- The 3D InceptionResNetV2 model achieved an AUROC of 0.916 on the external test set, outperforming radiologists and 2D image-based models.
- Radiologists' diagnostic accuracy improved by 16.9% (junior) and 11.1% (senior) with model assistance.
- 3D Grad-CAM visualizations confirmed the model focused on calcified regions, aligning with clinical diagnostic logic.

## Abstract

The presence of calcification, especially microcalcification, is often associated with an increased risk of malignancy and closely linked to papillary thyroid carcinoma (PTC), the most common type of thyroid cancer. However, existing diagnostic ultrasound (US) imaging has critical limitations such as inability to detect subtle calcifications via standard static imaging, leading to 15–20% delayed PTC treatment or unnecessary fine-needle aspiration. This study aimed to develop a calcification-optimized, interpretable deep learning (DL) model based on dynamic ultrasound videos to determine the malignancy nature of calcified thyroid nodules.

This study retrospectively collected ultrasound dynamic video data from 1,257 patients, containing 2,319 thyroid nodules across six hospitals between January 2020 and October 2023. Various DL models were constructed with optimization specifically implemented on the 3D InceptionResNetV2 network by including a calcification attention module to enhance sensitivity to micro-calcifications. Model performance was compared not only with those trained on 2D static ultrasound images, but also against diagnoses from four clinicians (2 junior and 2 senior radiologists). The dataset was split into training (70%, 1,623 videos), validation (10%, 232 videos), internal test (10%, 232 videos), and external test (10%, 232 videos) sets.

On the external test set, the optimized 3D InceptionResNetV2 model trained with dynamic videos outperformed the other four 3D DL models across all metrics: AUROC of 0.916, sensitivity of 0.860, and specificity of 0.834. Its AUROC was significantly higher than that of radiologists (0.916 versus 0.638; p < 0.0001). Additionally, with the assistance of the optimized model, radiologists’ diagnostic accuracy improved by 16.9% (junior) and 11.1% (senior) in the external cohort. 3D Grad-CAM further confirmed the model focused on calcified regions (consistent with clinical diagnostic logic) by generating interpretable heatmaps.

A calcification-optimized DL model trained on dynamic ultrasound videos was proposed to efficiently and accurately predict the benign/malignant nature of calcified nodules. This tool shows promises as a non-invasive, interpretable tool for early PTC detection, supporting timely diagnosis and treatment planning.

• This study developed a 3D InceptionResNetV2 deep learning model trained with dynamic ultrasound videos to identify and categorize calcified thyroid nodules.

• The 3D InceptionResNetV2 model was trained and tested using multi-center ultrasound video data from 1,257 patients to maximize its generalization ability in predicting the nature of calcified thyroid nodules.

• The 3D InceptionResNetV2 model, with an AUROC of 0.916, outperformed not only radiologists, but also models trained on the sole basis of static 2D images and clinical diagnoses.