波谱学杂志 ›› 2023, Vol. 40 ›› Issue (3): 293-306.doi: 10.11938/cjmr20223045

• 研究论文 • 上一篇    下一篇

融合注意力机制Mask RCNN的桥小脑角区听神经瘤和脑膜瘤的识别研究

胡小洋1,2,刘颖1,*(),陈淑2,董彬彬3   

  1. 1.上海理工大学 健康科学与工程学院,上海 200093
    2.上海伽玛医院放疗科,上海 200235
    3.上海伽玛医院放射科,上海 200235
  • 收稿日期:2022-12-15 出版日期:2023-09-05 在线发表日期:2023-03-21
  • 通讯作者: *Tel: 18602168660, E-mail: ling2431@163.com.
  • 基金资助:
    上海介入医疗器械工程技术研究中心(18DZ2250900)

Fusing Attention Mechanism with Mask RCNN for Recognition of Acoustic Neuroma and Meningioma in Cerebellopontine Angle

HU Xiaoyang1,2,LIU Ying1,*(),CHEN Shu2,DONG Binbin3   

  1. 1. School of Health Science and Engineering, University of Shanghai for Science and Technology, Shanghai 200093, China
    2. Department of Radiotherapy, Shanghai Gamma Hospital, Shanghai 200235, China
    3. Depart of Radiology, Shanghai Gamma Hospital, Shanghai 200235, China
  • Received:2022-12-15 Published:2023-09-05 Online:2023-03-21
  • Contact: *Tel: 18602168660, E-mail: ling2431@163.com.

摘要:

为探讨采用T1WI增强图像,利用融合注意力机制的掩膜区域神经网络(Mask RCNN)模型实现对桥小脑角区听神经瘤和脑膜瘤的识别.本文回顾性收集经病理或临床诊断确诊的脑膜瘤116例和听神经瘤427例,经图像筛选后共采用脑膜瘤872张和听神经瘤2 467张.按近似7:1.5:1.5的比例分为训练集、验证集和测试集.对图像进行预处理后,采用以Resnet50、Resnet101和VGG19为主干网络的Mask RCNN模型,以及融合卷积注意力机制的Mask RCNN模型Resnet101-CBAM和VGG19-CBAM对桥小脑角区听神经瘤和脑膜瘤进行检测和病灶分割.并使用均值平均精度(mean average precision,mAP)和均值平均召回率(mean average recall,mAR)评价模型性能.测试集结果显示卷积注意力机制可以提升模型性能,VGG19-CBAM模型在5个模型中综合性能最高,在分类和病灶分割的mAP分别为0.932和0.930.这表明融合注意力机制的Mask RCNN模型对桥小脑角区听神经瘤和脑膜瘤的识别较为理想,可为诊断和靶区勾画提供参考,提高临床工作效率.

关键词: 掩膜区域神经网络, 卷积注意力模块, 听神经瘤, 脑膜瘤, 磁共振图像

Abstract:

To investigate the performance of Mask RCNN (region-based convolutional neural network) using T1WI-enhanced images and fusing with attention mechanism in recognizing acoustic neuroma and meningioma in cerebellopontine angle area, this paper retrospectively collected 116 cases of meningioma and 427 cases of acoustic neuroma confirmed by pathology or clinical diagnosis. 872 images of meningioma and 2 467 images of acoustic neuroma were adopted after image screening. These images were divided into the training set, validation set, and test set in a ratio of approximately 7:1.5:1.5. After preprocessing the images, we employed five models, namely Mask RCNN model with Resnet50, Resnet101, VGG19 as the backbone network, and Mask RCNN models Resnet101-CBAM and VGG19-CBAM fusing with the convolution attention mechanism, for the detection and lesion segmentation of acoustic neuromas and meningioma in cerebellopontine angle area. The model performance was evaluated with mean average precision (mAP) and mean average recall (mAR). The results in the test set showed that the convolution attention mechanism could improve model performance. VGG19-CBAM model outperformed other models with mAP of 0.932 in classification and 0.930 in segmentation. This indicates that Mask RCNN fusing with attention mechanism has a better performance in recognizing acoustic neuroma and meningioma in cerebellopontine angle area, and hence can improve clinical efficiency by providing a reference for the diagnosis and target area segmentation.

Key words: Mask RCNN, CBAM, acoustic neuroma, meningioma, magnetic resonance image

中图分类号: