Brain Tumor Segmentation using Automatic 3D Multi-channel Feature Selection Convolutional Neural Network

Abstract

Brain tumor segmentation is an important topic in medical image processing. Three-dimensional (3D) convolutional neural networks (CNN) make full use of feature information and have better segmentation performance. However, conventional 3D CNNs involve a significant number of parameters and has high hardware requirements, which is not conducive to clinical application. To solve these issues, this study proposes a lightweight highly-efficient 3D CNN for brain tumor magnetic resonance imaging segmentation. We first replaced the conventional CNN with improved depth-wise CNN to save computational resources. Next, we employed a multi-channel convolution approach using convolution kernels of different sizes to aggregate features from different receptive fields. Finally, we used a squeeze-and-excitation unit to automatically extract useful feature information in the network. Application of the proposed model to the BraTS2019 validation set resulted in mean Dice scores of 0.9053, 0.8373, and 0.7847, with regard to whole tumor, tumor core, and enhanced tumor, respectively. This was achieved with 2.01 M parameters and 24.68 G floating point operations, and data training used 2.75 GB of video memory. The differences between our Dice scores and those derived using the network developed by the winner of the BraTS2019 challenge were only 0.0041, 0.0174, and 0.0274. Furthermore, the number of parameters and floating point operations were 2.3 times and 8.1 times lower and data training used 4 times lower video memory than the winning network. This indicates that the proposed network offers a higher precision for brain tumor segmentation while reducing computational requirements and effectively lowering hardware requirements for clinical applications.

Overview of convolutional neural networks architectures for brain tumor segmentation

“…Shan et al [48] proposed 3D CNN based on U-net architecture. Their model comprised three main units: improved depth-wise convolution (IDWC) unit which uses separable convolution instead of conventional convolution to extract feature maps and computationally saving resources.…”

A Feasibility Study on Deep Learning Based Brain Tumor Segmentation Using 2D Ellipse Box Areas

“…Das et al [ 13 ] used 3D CNN in a cascaded format to extract whole tumors first in a series followed by the core tumor and then the enhanced core tumor. Shan et al [ 14 ] proposed a lightweight 3D CNN with improved depth and used multi-channel convolution kernels of different sizes to aggregate features. Ramin et al [ 15 ] used a cascade CNN to speed up the learning.…”