Interpretation of cytomorphological features in thyroid fine needle aspiration biopsies using artificial intelligence

Thyroid nodules are a common clinical issue. The prevalence of palpable thyroid nodules is 1% in men and 5% in women worldwide, considering areas without iodine deficiency. After evaluation with clinical, laboratory, and ultrasonographic findings, diagnostic Fine Needle Aspiration Biopsy (FNA) may be performed if necessary. Fine needle aspiration biopsies play a significant role in the clinical management of thyroid nodules. The evaluation of FNA involves numerous criteria, and in daily routine practice, it can be challenging in terms of time and workload to have many preparations for the same case scanned under a light microscope by a pathologist for an extended period, and to examine the detailed cell groups observed. Additionally, the evaluation of aspiration samples requires special expertise, involving training and experience. This study aims to retrospectively select, photograph, and label FNA smears that have received a diagnosis in our department. Subsequently, with the help of a convolutional neural network model, an artificial intelligence program is intended to learn and classify the diagnostic concepts of 'benign' and 'malignant.' Our study includes 50 benign and 50 malignant diagnosed conventional smear preparations. After photographing the appropriate areas in all preparations, 721 images diagnosed as 'benign' and 710 images diagnosed as 'malignant' were obtained. Annotation, image augmentation, dataset creation, model architecture creation, model training, and statistical analysis were performed. According to the results obtained, the sensitivity rate determined for the benign diagnosis is 95.9%, the specificity rate is 95.6%, and the F1 score is determined as 0.958. For the malignant diagnosis, the sensitivity rate is 95.6%, the specificity rate is 95.9%, and the F1 score is determined as 0.957. The precision value calculated for both diagnosis groups is determined as 0.958. The Receiver Operating Characteristics (ROC) curve was plotted, and the area under the curve was calculated. When measured separately for benign and malignant diagnostic classes, it was found to be 0.99 for both. Our study has demonstrated that a successful model with similar sensitivity, specificity, F1 score, and area under the ROC curve statistics as other studies in the same field can be trained and adapted. The standardization of diagnostic categories, augmentation of input data through various steps, adjustment of model layer numbers and processes, and optimization of the training iterations led to the creation of a model with high statistical success. Among the limitations of our study and considerations for future research are the inclusion of preparations prepared using different staining and spreading techniques, working with regions of interest (ROI) selected autonomously by the model on Whole Slide Images (WSI), classification of all TSRBS categories, inclusion of patient history, clinical, and radiological features in input data, and the selection of cases where cytological diagnoses have been histologically confirmed. In the age of technology, the integration of artificial intelligence applications into our daily routines , especially in the field of medicine, will become one of the most essential aids for physicians in the diagnosis and management of various health issues.