Automatic Grading of Meningiomas Based on Residual Network Combined with Normalized Attention

Meningiomas are the most common primary intracranial tumors in adults, with varying degrees of aggressiveness, but the radiological presentation of high-grade meningiomas (HGM) and low-grade meningiomas (LGM) largely overlap, making accurate grading of meningiomas crucial. This study improves the performance of the model in grading meningiomas by suppressing the insignificant features of meningioma MRI images in convolutional neural networks. T1-enhanced weighted images of 180 patients with pathologically confirmed meningiomas were analysed. The dataset contained 58 high-grade meningiomas (HGM), (World Health Organization (WHO) grade II meningiomas and WHO grade III meningiomas), and 122 grade I meningiomas as low-grade meningiomas (LGM). To improve the grading accuracy of meningiomas, we added a normalization-based attention module (NAM) to ResNet, after which the features were mapped to high-grade meningiomas (HGM) and low-grade meningiomas (LGM) via two fully connected layers. In the training phase we used only raw images from normalized and non-degraded non-tumor regions as well as data augmentation techniques, as any augmentation method is very different from clinically used MRI images. In the inference phase, we evaluated the performance of our method in terms of accuracy, sensitivity, specificity and area under the curve (AUC). The conclusions show that T1-enhanced weighted images contain graded features of meningiomas and that deep learning methods can be used to clinically assess meningioma grading.