Application specific convolutional neural networks for brain tumor detection

Abstract

The research on CNN applications for medical image processing has been progressing rapidly. In the process of the development, hurdles appear and are to be overcome. The limitation in training samples is one of them, and restriction in computation resources can be another. In this paper, we present a design approach of application specific CNN (ASCNN), allowing to minimize the computational complexity of CNN systems without lowering the performance. This approach is to full-custom design CNNs for specific applications, such as brain tumor detection, so that each part of a CNN can be optimized to suit the input data and the task assigned to it. The convolution kernels and layers are made just-sufficient, nothing excessive. In this way, the randomness and the redundancy in computation can be minimized, the dependency on training samples decreased, the information density in data flow increased, the computation efficiency/quality and performance reliability improved. Three ASCNN systems for brain tumor detection are also presented as design examples. The results of the performance evaluation demonstrate that each of them delivers a high-quality detection with a computation volume of one-digit percentage, or less, of that needed by other CNN systems recently reported in reputed journals in the research area. Hence, ASCNN approach is effective to achieve high process quality at low computation cost. It can also lower the barrier of resource requirement of CNN systems to make them more implementable and applicable for general public.