Clinical knowledge embedded method based on multi-task learning for thyroid nodule classification with ultrasound...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Thyroid nodules are common, requiring accurate benign/malignant diagnosis for treatment.
• Current ultrasonography methods are subjective and variable.
• Deep learning models often overlook crucial clinical knowledge.

Purpose of the Study:

• To develop an accurate and reliable thyroid nodule classification model.
• To integrate clinical knowledge from the ACR Thyroid Imaging, Reporting and Data System guideline.
• To improve diagnostic performance by leveraging correlations between clinical features and malignancy.

Main Methods:

• A multi-task learning model was designed, incorporating ACR guideline features as tasks alongside pathological type.
• Clinical features were modeled to exploit correlations for improved diagnostic performance.
• A loss-weighting strategy was used to mitigate noisy label impacts.
• Model evaluation involved five-fold cross-validation on an internal dataset and testing on an external dataset.

Main Results:

• The multi-task learning model achieved an average AUC of 0.901.
• An ensemble approach further improved performance to an AUC of 0.917.
• These results significantly surpassed single-task baseline models.

Conclusions:

• Multi-task learning integrating clinical features effectively classifies thyroid nodules.
• Clinical indicators can serve as auxiliary tasks to enhance diagnostic performance for various diseases.
• The proposed method offers a more accurate and reliable approach to thyroid nodule diagnosis.