Brain Tumor Image Classification by Randomly Wired Neural Networks with a Modified Method

• Published in: 2020 IEEE 9th Data Driven Control and Learning Systems Conference (DDCLS) pp. 1141 - 1146
• Main Authors: Du, Xiaohao, Chen, Liangming, Liu, Zhiyi, Li, Shuai, Liu, Mei, Yang, Jun, Jin, Long
• Format: Conference Proceeding
• Language: English
• Published: IEEE 20.11.2020
• Subjects: A modified method; Biological neural networks; Brain modeling; Brain tumor diagnosis; Convolutional neural networks; Feature extraction; Image classification; Randomly wired neural networks (RWNN); Testing; Training; Tumors
• DOI: 10.1109/DDCLS49620.2020.9275242

Brain tumors including meningioma, glioma, and pituitary tumor are common tumors in the middle-aged and elderly people, which are not easy to be diagnosed and treated, thus effecting people's physical health and even life. In addition to the diagnosis of brain tumors by doctors based on the clinical experience and biochemical reaction results, computer-assisted diagnosis can also be relied upon, which is a product of the development of artificial intelligence. In this paper, a convolutional neural network with a randomly generated network structures is used to realize the diagnosis of brain tumor types by the magnetic resonance imaging (MRI) images classification. The presented neural network is called randomly wired neural network (RWNN). In addition, we propose a modified method for the RWNN model, which improves the accuracy of the RWNN model in image classification by about 1% without increased training time. Comparison experiments include comparing the modified RWNN model with the original RWNN model, the models proposed by other work and some classic convolutional neural network models such as ResNet and EfficientNet for image classification. Compared with other models, the modified RWNN model achieves the highest brain tumor image classification accuracy of 95.33%. These remarkable experimental results show that, the modified RWNN model is an effective and feasible tool for the diagnosis of brain tumors. In addition, it also provides a new research direction for the neural network structure design.