The place of magnetic resonance imaging in the discrimination of benign, borderline and malign in over lesions andcomparison with MR-based artificial intelligence

Objective:Ovarian cancer is the second most common cause of death in women gynecological cancers.Clinical findings are not specific and this causes delay in diagnosis.MRI, which is the most effective non-invasive method used in diagnosis today, has limitations, especially in distinguishing borderline and malignant masses. In this study, both radiomic features of ovarian mass images were extracted and a deep learning algorithm was developed by applying CNN architectures to distinguish benign borderline and malignant ovarian masses.The aim of this study is to determine and compare the accuracy rates of routine MRI, diffusion MRI sequences and MRI-based artificial intelligence applications in the differentiation of benign, borderline and malignant ovarian masses. Methods:This descriptive cross-sectional study was conducted with 191 patients who underwent contrast-enhanced lower abdomen and diffusion MR in a university hospital between April 1, 2015 and March 01, 2021.The ages and blood results of the patients were noted. Among the MRI findings, T1 and T2 signal characteristics, contrast enhancement patterns, presence of intra-mass hemorrhage, presence of mural nodules and papillary projections, presence or absence of diffusion restriction, ADC (x 10-3 mm2/s) values and radiological preliminary diagnoses were recorded as benign, borderline and malignant. With the Slicer 3D program, the masses were manually segmented in the axial T2 sequence, and a total of 107 features were extracted with the radiomics module.JPEG images of the sections in which masses were seen in axial T2 and contrast enhanced T1 sequences were taken and taught to CNN models (GoogleNet, ResNet18, MobileNetv2, ShuffleNet, VGG16, AlexNet and DarkNet19) as 80% training, 20% test and 90% training, 10% test group. The success of the developed CNN architectures in diagnosis was evaluated and the results were compared. To measure the success of the technique used in artificial intelligence methods, precision, sensitivity, specificity and accuracy success criteria were used. SPSS 20 program was used for statistical analysis and p<0.05 was considered significant. Results:Sensitivity 0.95, specificity 0.85, accuracy 0.78 and 0.80 precision for postcontrast T1 sequence in CNN architectures; Sensitivity for T2 sequence was 0.91, specificity 0.90, accuracy 0.84, and precision up to 0.90.Radiological findings showed 89.47% sensitivity, 87.31% specificity, 87.9% accuracy and 75% precision in diagnosing malignant mass. Conclusion: ESA architectures provided diagnostic performance not inferior to radiological findings in diagnosing ovarian cancer. Keywords:Ovarian cancer, Convolutional Neural Networks, Radiomics, MR