Deep learning model developed using multiparametric magnetic resonance imaging in the differential diagnosis of parotid gland tumors

Aim: To create a new artificial intelligence-based approach by developing a deep learning model using multiparametric MRI in the differential diagnosis of common parotid tumors. Material and Methods: MRI images of 123 patients diagnosed with parotid tumors and 50 patients without parotid tumors were analyzed retrospectively. With these images, parotid tumors were classified using the InceptionResNetV2 deep learning (DL) model and majority voting approach. The differential diagnostic performance of the model was evaluated using SVM, KNN, and LD classifiers. As input data in the DL model, post-contrast T1A, T2A conventional MRI sequences, and b0, b1000, and ADC diffusion-weighted sequences were used. The study was conducted in three stages. In Stage I, the differential diagnosis of control group (CG), pleomorphic adenoma (PMA), Warthin tumor (WT), malignant tumor (MT) was examined and there were two approaches in which MRI sequences were given as combined and non-combined in the input data. In Stage II; differential diagnosis of the benign tumor (PMA and WT), MT, and CG was made. In Stage III, it was aimed to make the differential diagnosis of patients with a tumor in the parotid gland and healthy parotid gland tissue. Results: In Stage I, the accuracy, sensitivity, and specificity in the non-combined approach were 86,43%, 83,24%, and 95,35%; in the combined approach, it was found to be 92,86%, 90,34%, and 97,51%, respectively. In Stage I, the SVM classifier had the best result among the classifiers with an accuracy of 93,57%. In Stage II, the accuracy was 92,14%, the sensitivity was 83,33% and the specificity was 93,99%. In Stage III, the accuracy was 99,29%, the sensitivity was 100%, and the specificity was 99,02%. The accuracy of the DL model for differential diagnosis of PMA was calculated 97,62%, WT 92,31%, and MT 71,43%. Conclusions: In this study, a new approach is presented that classifies common parotid tumors automatically and with high accuracy using DL models that have become popular in recent years, and statistically successful results have been obtained. This model will provide convenience for physicians in evaluating MRI images of parotid masses. Keywords: Artificial Intelligence, Deep Learning, Parotid Tumors, Computer Aided Diagnosis