Knowledge-guided synthetic medical image adversarial augmentation for ultrasonography thyroid nodule classification

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning excels in image classification but requires large annotated datasets, which are scarce in the medical domain.
• Acquiring large-scale, annotated medical datasets remains a significant challenge for developing robust AI models.

Purpose of the Study:

• To propose a knowledge-guided adversarial augmentation method for synthesizing high-quality medical images.
• To address the data scarcity issue in medical AI by leveraging expert knowledge.

Main Methods:

• Developed Term and Image Encoders to extract domain knowledge from radiologists.
• Constrained an Auxiliary Classifier Generative Adversarial Network (ACGAN) with extracted domain knowledge for image synthesis.
• Applied the method to synthesize thyroid nodule images and classify them using ultrasonography data.

Main Results:

• Achieved 91.46% accuracy, 90.63% sensitivity, 92.65% specificity, and 95.32% AUC for thyroid nodule classification on a limited dataset.
• Demonstrated superior performance compared to existing thyroid nodule classification methods.
• Showcased enhanced generalization and robustness of the proposed model.

Conclusions:

• The proposed method effectively alleviates data insufficiency in medical AI applications.
• Synthetic augmentation using this knowledge-guided approach holds promise for various medical problems.
• The method leverages expert diagnostic experience extracted from standardized terms.