Automatic Precise Segmentation of Cerebellopontine Angle Tumor Based on Faster-RCNN and Level-Set Method

Abstract

To meet the demands in surgical treatment and radiotherapy, this work combines the faster region convolutional neural network (Faster-RCNN) and Level-Set methods to segment cerebellopontine angle (CPA) tumors automatically and precisely. T1WI-SE magnetic resonance images from 317 CPA tumor patients were collected. Features extracted by VGG16 were combined with the region proposal network (RPN) for training. A CPA tumor localization model was then established, before the Level-Set method was applied to accurately segment the tumor. The segmentation results of different CPA tumor regions were compared in terms of precision, recall, mean average precision (mAP) and Dice coefficient. The results showed that the method proposed can effectively and precisely segment CPA tumors, thereby capable of reducing the burden on clinicians and improving the treatment effect.

Keywords

• faster region convolutional neural network (faster-rcnn)
• level-set
• image segmentation
• cerebellopontine angle tumor