AS-3DFCN: Automatically Seeking 3DFCN-Based Brain Tumor Segmentation

Convolutional neural network, an optimization technique inspired by biological visual cortex, is widely used in pattern recognition, information processing, and other fields, and it has a good ability to solve nonlinear problems. However, there are some disadvantages, such as its uninterpretability, high computational cost, and GPU consumption, which make its structural design difficult. As a result, the existing neural networks require manual design by experts, and numerous experiments have been conducted to assess their effectiveness. In order to reduce the difficulty of designing neural networks manually, and to design more effective networks to solve problems, we propose a model for automatically seeking 3D fully convolutional network (AS-3DFCN) based on genetic algorithm. Firstly, we suggest an AS-3DFCN that defines the search space as a directed graph. Secondly, in order to improve the segmentation accuracy of our model for brain tumor images, the genetic algorithm is applied to search an ideal topology network. Finally, we evaluate the effectiveness of our proposed algorithm using the MICCAI-BraTS2018-2020 public datasets. After comparing with other manually designed fully convolutional networks, our method outperforms many other advanced methods, with average Dice coefficients of 84.7, 83.1, and 83.3; average Hausdorff measures of 4.32, 4.88, and 13.5; and parameter quantities of only 5.34M. It has also shown that our model has fewer parameters, can automatically seek the optimum network topology, and can segment brain tumor images more accurately.