Identification method of thyroid nodule ultrasonography based on self-supervised learning dual-branch attention...

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Radiology and Diagnostic Imaging

Background:

• Thyroid ultrasound is crucial for nodule detection but faces challenges due to low image contrast, noise, and heterogeneity.
• Limited availability of high-quality labeled medical imaging datasets hinders machine learning applications in thyroid ultrasound analysis.

Purpose of the Study:

• To propose a novel Dual-branch Attention Learning (DBAL) convolutional neural network framework for enhanced thyroid nodule detection.
• To improve the generalization ability of machine learning models using limited data through jigsaw puzzle pretext tasks.
• To accurately discriminate between malignant and benign thyroid nodules using AI-driven image analysis.

Main Methods:

• Developed a Dual-branch Attention Learning (DBAL) convolutional neural network framework to capture contextual information in thyroid ultrasound images.
• Employed a jigsaw puzzle pretext task during network training to enhance generalization with limited data.
• Utilized self-supervised pre-training on unlabeled ultrasound images followed by fine-tuning on 1216 clinical ultrasound images.

Main Results:

• The DBAL framework demonstrated effective capture of intrinsic features in a global-to-local manner.
• Achieved an 88.5% correct diagnosis rate for differentiating malignant and benign thyroid nodules.
• Obtained a 93.7% area under the ROC curve, indicating high diagnostic performance.

Conclusions:

• The proposed DBAL framework shows significant accuracy and efficiency in thyroid nodule detection and classification.
• The approach effectively addresses the challenges of limited labeled data in medical imaging through innovative pre-training strategies.
• DBAL holds promising potential for clinical application in improving the diagnosis of thyroid nodules.