Preliminary Results of Deep Learning Approach for Preoperative Diagnosis of Ovarian Cancer Based on Pelvic MRI Scans.

Abstract

To predict the pathological diagnosis of ovarian tumors using preoperative MRI images, using deep learning models. A total of 185 patients were enrolled, including 40 with ovarian cancers, 25 with borderline malignant tumors, and 120 with benign tumors. Using sagittal and horizontal T2-weighted images (T2WI), we constructed the pre-trained convolutional neural networks to predict pathological diagnoses. The performance of the model was assessed by precision, recall, and F1-score on macro-average with 95% confidence interval (95%CI). The accuracy and area under the curve (AUC) were also assessed after binary transformation by the division into benign and non-benign groups. The macro-average accuracy in the three-class classification was 0.523 (95%CI=0.504-0.544) for sagittal images and 0.426 (95%CI=0.404-0.446) for horizontal images. The model achieved a precision of 0.63 (95%CI=0.61-0.66), recall of 0.75 (95%CI=0.72-0.78), and F1 score of 0.69 (95%CI=0.67-0.71) for benign tumor. Regarding the discrimination between benign and non-benign tumors, the accuracy in the binary-class classification was 0.628 (95%CI=0.592-0.662) for sagittal images and AUC was 0.529 (95%CI=0.500-0.557). Using deep learning, we could perform pathological diagnosis from preoperative MRI images.