[Auxiliary pathological diagnosis algorithm based on color moments for frozen-section of thyroid cancer]

Objective: To develop a color-moment based model for frozen-section diagnosis of thyroid lesions, and to evaluate the model's value in the frozen-section diagnosis of thyroid cancer. Methods: In this study, 550 frozen thyroid pathological slides, including malignant and non-malignant cases, were collected from Taizhou Central Hospital (Taizhou University Hospital), China, between June 2018 and January 2020. The 550 digitalized frozen-section slides of thyroid were divided into training set (190 slides), validation set (48 slides), test set A (60 slides) and test set B (252 slides). The tumor regions on the slides of malignant cases in the training and validation sets were labeled by pathologists. The labeling information was then used to train the thyroid frozen-section diagnosis models based on the voting method and those based on the color moment. Finally, the performance of two pathological slide diagnosis models was evaluated using the test set A and test set B, respectively. Result: The classification accuracy of the thyroid frozen-section diagnosis model based on the voting method was 90.0% and 83.7%, using test sets A and B, respectively, while that based on color moments was 91.6% and 90.9%, respectively. For actual frozen-section diagnosis of thyroid cancer, the model developed in this study had higher accuracy and stability. Conclusion: This study proposes a color-moment based frozen-section diagnosis model, which is more accurate than other classification models for frozen-section diagnoses of thyroid cancer.