Objective: to evaluate the effectiveness of a convolutional neural network model for automated cytologic diagnosis of papillary thyroid cancer and benign thyroid nodules.Material and Methods. The convolutional neural network was developed in the Python programming language using the TensorFlow 2.15.0 open source library. For the study, a dataset that included two categories of pathologies was generated: 1597 microphotographs of papillary carcinoma and 767 microphotographs of benign nodules (colloid goiter and adenomatous nodules). To form a training sample and evaluate the model’s performance metrics on the test sample, the dataset was divided in a ratio of 80/20.Results. In classifying papillary carcinoma, the model achieved precision of 89.3 %, recall of 92.4 %, specifcity of 77.4 % and F1 score of 91.4 %. When identifying benign nodules, the presicion, recall, specifcity and F1 score were 83.3 %, 77.4 %, 92.4 %, and 80.3 %, respectively, indicating a higher rate of false-positive and false-negative predictions. The AUC was 0.91 at the individual microphotograph level and 0.94 at the serial microphotograph level from one patient, indicating the high ability of the trained model to differentiate between malignant and benign thyroid lesions based on microphotographs of fne-needle aspiration biopsy specimens.Conclusion. Further improvement of the neural network model by training on larger and more diverse datasets of microphotographs of cytological specimens of the thyroid gland will help improve its diagnostic range and performance. The developed model can be used to develop software for identifying thyroid pathologies.
Read full abstract