Deep convolutional neural network model ResNeSt for discrimination of papillary thyroid carcinomas and benign nodules in thyroid nodules diagnosed as atypia of undetermined significance.

Background: A deep convolutional neural network (DCNN) model was employed for the differentiation of thyroid nodules diagnosed as atypia of undetermined significance (AUS) according to the 2023 Bethesda System for Reporting Thyroid Cytopathology (TBSRTC). The aim of this study was to investigate the efficiency of ResNeSt in improving the diagnostic accuracy of fine-needle aspiration (FNA) biopsy. Methods: Fragmented images were used to train and test DCNN models. A training dataset was built from 1,330 samples diagnosed as papillary thyroid carcinoma (PTC) or benign nodules, and a test dataset was built from 173 samples diagnosed as AUS. ResNeSt was trained and tested to provide a differentiation. With regard to AUS samples, the characteristics of the cell nuclei were compared using the Wilcoxon test. Results: The ResNeSt model achieved an accuracy of 92.49% (160/173) on fragmented images and 84.78% (39/46) from a patient wise viewpoint in discrimination of PTC and benign nodules in AUS nodules. The sensitivity and specificity of ResNeSt model were 95.79% and 88.46%. The κ value between ResNeSt and the pathological results was 0.847 (P<0.001). With regard to the cell nuclei of AUS nodules, both area and perimeter of malignant nodules were larger than those of benign ones, which were 2,340.00 (1,769.00, 2,807.00) Conclusions: In summary, the DCNN model ResNeSt showed great potential in discriminating thyroid nodules diagnosed as AUS. Among those nodules, malignant nodules showed larger and more deeply stained nuclei than benign nodules.