Deep learning to assist composition classification and thyroid solid nodule diagnosis: a multicenter diagnostic...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• High-resolution ultrasound increases detection of thyroid nodules, leading to unnecessary fine-needle aspiration (FNA) and patient anxiety.
• Differentiating benign from malignant thyroid nodules is crucial for appropriate patient management.

Purpose of the Study:

• To develop a deep learning (DL) framework for classifying thyroid nodule composition and assessing malignancy risk.
• To evaluate the performance of convolutional neural network (CNN) models in differentiating benign from malignant solid thyroid nodules.

Main Methods:

• A retrospective multicenter study utilizing ultrasound images from 6784 nodules (11,201 images).
• CNN models, including Inception-ResNet, were trained and validated for nodule classification.
• Area under the receiver-operating characteristic curve (AUC) was the primary evaluation metric.

Main Results:

• CNN models achieved AUCs > 0.91 for solid thyroid nodule grading, with Inception-ResNet reaching 0.94.
• The best DL algorithm demonstrated a sensitivity of 0.88 and specificity of 0.86 in test sets.
• The DL model outperformed senior physicians (p < 0.001) in differentiating benign from malignant nodules.

Conclusions:

• CNN-based DL frameworks can effectively assist in thyroid nodule diagnosis.
• This technology has the potential to significantly reduce unnecessary fine-needle aspiration procedures.
• DL models offer a promising tool for improving the accuracy and efficiency of thyroid cancer risk assessment.