Rule-based automatic diagnosis of thyroid nodules from intraoperative frozen sections using deep learning

Area of Science:

• Pathology
• Computer Science
• Medical Imaging

Background:

• Intraoperative diagnosis of thyroid nodules from frozen sections is crucial for surgical guidance but presents challenges for pathologists.
• Existing methods may lack the speed and accuracy required for real-time surgical decision-making.

Purpose of the Study:

• To develop and validate a deep learning-based, rule-integrated system for automated differentiation of thyroid nodules in intraoperative frozen sections.
• To enhance diagnostic accuracy and efficiency, providing rapid, reliable results within the surgical workflow.

Main Methods:

• A three-component system was developed: automatic tissue localization in whole slide images (WSIs), patch classification using a fine-tuned InceptionV3 convolutional neural network (CNN), and a rule-based integration of patch predictions for final diagnosis.
• The CNN was trained to classify patches as benign, uncertain, or malignant.
• A rule-based protocol was designed for integrating patch-level predictions to achieve a final slide diagnosis with interpretability.

Main Results:

• The system achieved high accuracy, correctly predicting 95.3% of benign and 96.7% of malignant nodules on 259 testing slides.
• 16.2% of slides were classified as uncertain, requiring further pathologist review.
• Diagnosis of a typical whole slide image was completed within 1 minute, meeting intraoperative time constraints.

Conclusions:

• The proposed deep learning and rule-based system demonstrates significant potential for accurate and rapid intraoperative diagnosis of thyroid nodules from frozen sections.
• This approach offers a valuable tool to assist pathologists, improve diagnostic efficiency, and support surgical decision-making.
• This represents the first application of deep learning for diagnosing thyroid nodules on intraoperative frozen sections.