Classification and Localization of Meningioma and Acoustic Neuroma in Cerebellopontine Angle Based on Mask RCNN

Abstract

Differential diagnosis of meningioma and acoustic neuroma can be difficult because these two tumors have similar locations and appearances on medical images. To address this problem, mask region convolutional neural network (Mask RCNN) was used to classify and diagnose those two types of tumors. First, magnetic resonance images acquired with T1-weighted spin-echo (T1WI-SE) sequence of 89 meningioma and 218 acoustic neuroma patients were collected and preprocessed. Then the improved feature pyramid networks (FPN) algorithm was used for model network training. The effects of three different backbone feature extraction layers on classification and location were compared. It was demonstrated that Mask RCNN model with improved FPN and ResNet101 as backbone network is able to effectively classify and locate meningioma and acoustic neuroma, the values of precision, recall, specificity, and mean average precision (mAP) are 0.918 2, 0.856 9, 0.876 2, and 0.90, respectively.

Keywords

• mask rcnn
• fpn algorithm
• classification and localization
• meningioma
• acoustic neuroma