Can deep learning replace fine needle aspiration biyopsy(FNAB) indi̇agnosi̇s of benign and malign parotid masses?

ABSTRACT Objective and Aim: Parotid tumors are tumors that can be found in all age groups and have many different histological subtypes. Parotid glands are the source of 80% of all salivary gland tumors and curative primary treatments are usually surgery. However, the scope of surgery to be performed depends on whether the tumor is benign or malignant. An attempt is made to make a histopathological diagnosis through preoperative invasive interventions. Fine needle aspiration biopsy (FNAB) and surgical biopsy procedures have led to the search for economical and comfortable new methods that can provide less risky and more accurate results due to the risks of complications, difficulties in patient compliance, and loss of time and labor experienced before and after these procedures. All Artificial intelligence applications, which are rapidly developing in the world and started to be used in many areas of our lives, have also started to be tested in the diagnosis and treatment of medical diseases. In this study, we aimed to evaluate the ability of the ResNet50 architecture, one of the deep learning models, which is a sub-title of machine learning, one of the artificial intelligence applications, to distinguish normal parotid tissue, benign and malignant parotid masses, with reference to pathology results and to compare it with previous studies on this subject. Materials and Methods: Our study was conducted retrospectively by selecting patients who applied to Dicle University Faculty of Medicine Ear Nose and Throat clinic with a parotid mass between 01.01.2014 and 01.07.2024. Two groups diagnosed with benign and malignant tumors were created, as well as a group of normal patients whose contrast-enhanced neck CT was performed for any purpose for the training of the deep learning model and no pathology was detected in the parotids. 52 patients in each group, 156 patients in total, were included in the study. From the axial section CT images of each patient, 6 images passing through the parotid and mass levels were taken. The dataset consisting of 936 images in total was used as input in the ResNet50 architecture. 70% of this data set was used for training, 30% for testing, and 10% of the data set used in training was used for validation. With the data obtained, training and validation accuracy, loss graphs, confusion matrix and model performance measurement results were determined. Results: In our study, 40.4% (n=21) of benign patients were female, 59.6% (n=31) were male, and 40.4% (n=21) of malignant patients were female, 59.6% (n=31) were female. =31) were male. While the average age in benign patients was 42.8, it was 62.2 in malignant patients. The average age was found to be statistically significantly higher in the malignant group. (p<0.001). The most common tumor in the benign group was pleomorphic adenoma with a rate of 65.4% (n=34). The most common tumor in the malignant group was mucoepidermoid carcinoma with 40% (n = 21). When the images were tested after training with the ResNet50 architecture, the positive predictive value (precision), sensitivity (recall) and f1-score were 96, 100, 98%, respectively, for the benign group, while they were 100, 93, 97%, respectively, for the malignant group. Conclusions: The deep learning model (ResNet50) we tested was found to be quite successful in both tumor types, and considering that it can be tested and developed in larger case series, and that the models to be developed for tumor subtypes and rapid developments in the field of artificial intelligence can enable these models to learn more successful features, we believe that FNAB studies will be performed in the near future. We think it can replace. Key words: parotid, tumor, deep learning, ResNet50, FNAB,