Medical Image Analysis With High Dimensional And Limited Training Samples

This paper is devoted to explore the method of medical image analysis with highdimensional limited training samples.In terms of limited training samples,we propose a bimodality medical image synthesis approach based on sequential generative adversarial network(GAN)and semi-supervised learning.Our approach consists of two generative modules that synthesize images of the two modalities in a sequential order.A method for measuring the synthesis complexity is proposed to automatically determine the synthesis order in our sequential GAN.Images of the modality with a lower complexity are synthesized first,and the counterparts with a higher complexity are generated later.Our sequential GAN is trained end-to-end in a semisupervised manner.In supervised training,the joint distribution of bi-modality images is learned from real paired images of the two modalities by explicitly minimizing the reconstruction losses between the real and synthetic images.To avoid overfitting limited training images,in unsupervised training the marginal distribution of each modality is learned based on unpaired images by minimizing the Wasserstein distance between the distributions of real and fake images.We comprehensively evaluate the proposed model using two synthesis tasks based on three types of evaluate metrics and user studies.Visual and quantitative results demonstrate the superiority of our method to the state-of-theart methods,and reasonable visual quality and clinical significance.For high-dimensional samples,this study presents an end-to-end trainable convolutional neural network(CNN)which combines three dimension(3D)CNN with two dimension(2D)CNN where the two steps are optimized jointly.The proposed CNN consists of three concatenated subnets: 1)a novel 3D candidate proposal network for detecting cubes containing suspected lesions;2)a 3D spatial transformation subnet for generating fixed-sized vessel-aligned 2.5D image representation for candidates,and 3)a 2D classification network eliminates fales positives.Compared with tranditional methods which typically employ separate steps,our methods have four advantages: 1)efficiently addressing the 3D spatial information;2)end-to-end learning for global optimization,and 3)3D CNN + 2D CNN for avoiding overfitting.We have evaluated our approach on the public dataset and locally collected dataset with FROC curve that demonstrates the superior performance of our model.