Ovarian masses suggested for MRI examination: assessment of deep learning models based on non-contrast-enhanced MRI sequences for predicting malignancy.

Abstract

Purpose: We aims to assessed and compare four deep learning(DL) models using non-contrast-enhanced magnetic resonance imaging(MRI) to differentiate benign from malignant ovarian tumors, considering diagnostic efficacy and associated development costs.

Methods: 526 patients (327 benign lesions vs 199 malignant lesions) who were recommended for MRI due to suspected ovarian masses, confirmed with histopathology, were included in this retrospective study. A training cohort (n=367) and a validation cohort (n=159) were constructed. Based on the images of T1-weighted imaging (T1WI), T2-weighted imaging (T2WI), diffusion-weighted imaging (DWI), we evaluated the diagnostic performance of four DL models (ConvNeXt, FBNet, GhostNet, ResNet50) in distinguishing between benign and malignant ovarian tumors. Two radiologists with varying levels of experience independently reviewed all original non-contrast-enhanced MR images from the validation cohort to determine if each case was benign or malignant. The area under the receiver operating characteristic curve (AUC), confusion matrices, accuracy, sensitivity, specificity, positive predictive value(PPV) and negative predictive value(NPV) were used to compare performance.

Results: The study of 526 ovarian mass patients (ages 1-92) evaluated four DL models for predicting malignant tumors, with AUCs ranging from 0.8091 to 0.8572 and accuracy between 81.1% and 85.5%. An experienced radiologist achieved 86.2% accuracy, slightly surpassing the DL models, while a less experienced radiologist had 69.2% accuracy. Resnet50 had the highest sensitivity (78.3%) and NPV (87.3%), while ConvNeXt excelled in specificity and PPV (100%). GhostNet and FBNet are more parameter-efficient than other models.

Conclusion: The four DL models effectively distinguished between benign and malignant ovarian tumors using non-contrast MRI. These models outperformed less experienced radiologists and were slightly less accurate than experienced ones. ResNet50 had the best predictive performance, while GhostNet was highly accurate with fewer parameters. Our study indicates that DL models based on non-contrast-enhanced MRI have the potential to assist in diagnosis.