Deep learning to diagnose Hashimoto's thyroiditis from sonographic images.

Area of Science:

• Endocrinology
• Medical Imaging
• Artificial Intelligence

Background:

• Hashimoto's thyroiditis (HT) is the primary cause of hypothyroidism, a common endocrine disorder.
• Accurate and timely diagnosis of HT is crucial for effective patient management and preventing complications.

Purpose of the Study:

• To develop and evaluate a deep learning model, HTNet, for the automated diagnosis of Hashimoto's thyroiditis using thyroid ultrasound images.
• To compare the diagnostic performance of HTNet against experienced radiologists.

Main Methods:

• A deep learning model (HTNet) was trained on a large dataset of 106,513 thyroid ultrasound images from 17,934 patients.
• The model's performance was validated on independent datasets comprising static images and video data from 5,051 patients.
• Integration of serologic markers with imaging data was explored to assess performance enhancement.

Main Results:

• HTNet achieved high diagnostic performance with an area under the receiver operating curve (AUC) of 0.905 on static images and 0.895 on video data.
• The model demonstrated superior accuracy (83.2% vs. 79.8%) and sensitivity (82.6% vs. 68.1%) compared to radiologists.
• Integrating serologic markers further improved HTNet's AUC to 0.949 for video data and 0.914 for static images.

Conclusions:

• HTNet shows significant potential as an automated tool for diagnosing Hashimoto's thyroiditis.
• The model's performance surpasses that of human radiologists, suggesting its utility in clinical settings.
• Combining imaging data with serologic markers can further enhance diagnostic accuracy for HT.