Segmentation of vestibular schwannoma from multi-parametric magnetic resonance images using convolutional neural network

In this study, we aim to automatically segment the Vestibular Schwannoma (VS) from multi-parametric magnetic resonance (MR) images before the Gamma Knife (GK) treatment using the deep learning based Convolutional Neural Network (CNN). 516 VS subjects' MR images and tumor contours were collected from Taipei Veteran General Hospital, Taiwan. All the MR images were scanned by 1.5 T GE scanner. The tumor contours were delineated manually by experienced neuroradiologists. MR images included 1) 1) T1- weighted (T1W) with matrix size 512 x 512, voxel size 0.5 x 0.5 x 3mm;2) T1- weighted gadolinium contrast-enhanced (T1W+C) with matrix size 512 x 512 and voxel size 0.5 x 0.5 x 3mm; 3) T2 - weighted (T2W) with matrix size 512 x 512, voxel size 0.5 x 0.5 x 3mm. Since some tumors consisted of solid part, which appeared as high intensity at T1W+C, and cystic part, which appeared as high intensity at T2W, we used multi-parametric MR images and designed a deep learning based encode-decode CNN model with two convolution pathways and different convolution kernel sizes at encode part to extract feature maps from different direction of anisotropic voxel-size MR images. Our results showed that the multi-parametric input, namely, T1W, T1W+C and T2W images, for the proposed CNN achieved superior performance with Dice coefficient = 0.87+0.06 in the segmentation of VS, especially for tumors with cystic components, compared to using the single-parametric input T1W+C image with Dice coefficient = 0.83+0.11.