Automatic Segmentation Of MRI Cardiac Images Based On Deep Learning

It is the most effective way to use a convolutional network to segment cardiac images,but it needs a lot of data.Reusing pre-trained parameters,such as semi-supervised and transfer learning,is one of the most important strategies to address the issue of data inadequacy.However,the fundamental reason for the success of these methods is still unclear.In this paper,we give an explanation and propose a solution that can not only evaluate whether a given network is reusable by means of the performance of reusing convolution kernels,but also evaluate which layers’ parameters of the given network can be reused by means of the performance of reusing corresponding parameters and,ultimately,evaluate whether those parameters are reusable in a target task by means of the root mean square error(RMSE)of the corresponding convolution kernels.Specifically,a CNN parameters can be reused,which depends upon two conditions:first,the network is a reusable network;and second,the RMSE is small enough between the convolution kernels from the source domain and target domain.We applied this method to cardiac segmentation,and experimental results demonstrate that the performance of using reusable parameters is significantly improved when meeting these conditions in target tasks.