Risk assessment of thyroid nodules with a multi-instance convolutional neural network.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Ultrasonography is the primary imaging method for thyroid nodules.
• Existing AI models for thyroid cancer diagnosis have limited performance.
• Need for improved AI-assisted diagnostic tools for thyroid nodules.

Purpose of the Study:

• To propose a novel multi-instance learning (MIL) convolutional neural network (CNN) model for ultrasound-based thyroid cancer diagnosis.
• To enhance the diagnostic performance and risk assessment of thyroid nodules using AI.
• To evaluate the clinical utility of the developed AI framework.

Main Methods:

• Developed a MIL-CNN model to extract nodule-level ultrasound features from instance-level images.
• Utilized an attention mechanism for feature aggregation and localization of key diagnostic indicators.
• Trained and tested the model on a dataset of 2000 ultrasound images from thyroid nodule patients.

Main Results:

• The proposed MIL-CNN model demonstrated superior performance compared to other classification models.
• Key performance metrics including accuracy, precision, recall, F1-Score, and AUC showed significant improvements.
• Statistical analysis confirmed the significant performance difference (p <0.0001).

Conclusions:

• The MIL-CNN framework is effective for non-invasive thyroid nodule risk stratification.
• The AI model supports informed clinical decisions, potentially reducing unnecessary biopsies and surgeries.
• The study highlights the clinical utility of advanced AI in thyroid cancer diagnostics.