Multi Objective Evolutionary Architecture Search Of Convolutional Network For Medical Image Segmentation

With the remarkable development of large-scale matrix parallel computing capability of graphics processor,the application of deep learning networks is rapidly spreading in various fields of computer vision.U-Net and its derivative networks have played important roles in the field of medical image semantic segmentation,which is credit to their excellent performances in deep semantic information extraction,feature transfer and information fusion between layers.When manually designing U-Net for different medical image segmentation tasks,it is necessary to select the hyper-parameters,such as the kernel size and the channels of convolution.And the search of architecture hyper-parameters is time-consuming.At the same time,due to the huge differences of computing resources and storage on mobile phones,cloud servers and other deployment devices,the scales of networks that can be deployed on different devices are different.It is difficult to apply a certain manually designed U-Net is to different devices.Therefore,this paper introduces neural architecture search(NAS)into U-Net for medical image segmentation,and designs super diamond atrous U-Net(diamond atrous U-Net,DAUNet),serving as the architecture space for subnet.Furthermore,an efficient multi-objective evolutionary neural architecture search algorithm for DAU-Net is proposed after considering the segmentation error and architecture parameters as the optimization objectives.In order to speed up NAS,weight sharing strategy and proxy model approximate evaluation strategy are designed.For a given medical image segmentation task,the proposed algorithm can efficiently search a group of excellent U-Nets with less parameters and low segmentation error,which is suitable for different deployment devices.The main contents of this paper are as follows.(1)In view of the defects of U-Net,such as the information loss of up and down sampling,the lack of deformation of receptive fields and parameters redundancy in deep layers,this paper proposes a diamond atrous convolution module DAB based on the residual network,inception structure and atrous convolution.Each convolution module in DAU-Net is a diamond atrous convolution module.The DAB combines three different atrous convolution operations,and uses the conventional 3×3 convolution for feature fusion.DAB can helps to expand the receptive field and reduce the amount of network parameters.On the basis of DAB,a lightweight channel attention mechanism is introduced into skip connection of DAU-Net,and the up and down sampling method are modified.Thus,a super DAU-Net is designed,which can speed up NAS by weight sharing and channel selection.(2)Based on DAU-Net,an efficient multi-objective evolutionary neural architecture search algorithm is designed.Firstly,the search space and chromosome coding scheme of subnet architecture are designed,and the search goals are to minimize segmentation error and architecture parameters.The convolution cores of DAU-Net subnet are consistent with those of supernet,but the number of convolution channels in each DAB module can vary within the maximum number of channels of supernet.In this paper,NSGA-III is used as the main framework of multi-objective evolutionary neural architecture search algorithm,and two kinds of architecture performance evaluation acceleration strategies are further designed.On the one hand,inspired by transfer learning,a weight sharing strategy based on supernet-subnet channel sorting and selection is designed,in which subnets can inherit the corresponding weights from the well-trained supernet according to the importance of the channels,thus reducing the time to evaluate performance of subnets.On the other hand,an approximate evaluation strategy based on random forest online proxy model is designed,which can efficiently predict the performances of future subnets by fully mining the historical data of subnets performance to further reduce the cost of evaluation time.The experimental results on ISIC2018 and DRIVE datasets show that compared with the medical image segmentation networks designed by other researchers recently,the prediction accuracy of DAU-Net achieve some improvement.This paper proposes an efficient multiobjective evolutionary neural architecture search algorithm,and the subnets on Pareto edge have obvious improvement compared with the manually designed network in terms of parameters and prediction accuracy.Moreover,for a given medical image segmentation task this algorithm can automatically search out a group of excellent U-Nets that take both parameters and segmentation accuracy into account,which are suitable for different deployment devices.