Multitask network for thyroid nodule diagnosis based on TI-RADS.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Thyroid nodule assessment relies heavily on radiologist expertise, making it time-consuming and subjective.
• Current deep learning models lack transparency, presenting a "black-box" challenge for clinical adoption.
• Explainable AI is crucial for integrating AI into clinical workflows and building physician trust.

Purpose of the Study:

• To develop an explainable deep learning model for thyroid nodule classification.
• To integrate the Thyroid Imaging Reporting and Data System (TI-RADS) into a multitask network for interpretable predictions.
• To provide objective recommendations that align with the clinical diagnostic process.

Main Methods:

• A multitask network based on TI-RADS (MTN-TI-RADS) was developed for thyroid nodule classification.
• The network utilizes multitask learning to obtain TI-RADS classifications, calculate points and risk levels, and classify nodules as benign or malignant.
• Attention modules were incorporated to enhance focus on critical image features.

Main Results:

• MTN-TI-RADS demonstrated superior performance compared to senior radiologists in classifying thyroid nodules.
• High sensitivity (0.988 internal, 0.949 external) and specificity (0.912 internal, 0.930 external) were achieved.
• The model achieved an area under the ROC curve of 0.981 (internal) and 0.973 (external), surpassing radiologists' scores.

Conclusions:

• MTN-TI-RADS is an effective, interpretable end-to-end network for thyroid nodule classification.
• The model offers superior classification ability and builds physician trust through its explainable nature.
• The network's alignment with clinical diagnosis processes indicates significant potential for clinical application.