TyroScan-AI software - convolutional neural network for binary classification of thyroid cytology preparations.
Increasing the speed and accuracy of automated diagnosis of papillary cancer and benign thyroid masses by analyzing cytological biopsy images using a convolutional neural network.
Biotechnology And National Health
Technologies for human health
6 months
Donetsk People's Republic
In the thyroid gland, 48% of patients are found to have nodules during routine physical examination, which may be benign or malignant. Benign nodules include colloid nodules, cysts, and inflammatory disease, while malignant nodules include various cancers such as papillary, follicular, and medullary. Cytologic examination by fine-needle aspiration puncture biopsy is the gold standard for preoperative diagnosis, but has high variability and may yield indeterminate results. Surgical resection of thyroid nodules can determine the nature of the neoplasm, but can be stressful and may lead to long-term adverse effects. To avoid unnecessary surgical interventions, fast and accurate methods for early diagnosis of thyroid cancer are needed, among which tools based on artificial neural networks that can process ultrasound, radiologic, and cytologic data are gaining popularity. The researchers of the laboratory of cell and tissue cultivation of the FSBI V.K. Gusak IERS of MOH of Russia developed a model of convolutional neural network in Python programming language using open source library TensorFlow 2.15.0. The dataset of the IERS's Endocrine Surgery Center (2364 images of papillary cancer, biggest in Russia) was used for training the neural network. In the classification of papillary carcinoma, the model achieved an accuracy of 89.3%, recall of 92.4%, and F1 score of 91.4%. In the classification of benign lesions, the accuracy was 83.3%, recall 77.4%, and F1 score 80.3%. The AUC score was 0.91 at the individual image level and 0.94 at the patient level, indicating the high ability of the trained model to differentiate between malignant and benign thyroid cytologic images. Based on this convolutional neural network, software (in Java Script programming language) TyroScan-AI was developed and patented. This software was introduced into clinical practice at the Endocrine Surgery Center to accelerate screening studies of thyroid biopsies.
N.A.
N.A.
