Medical Image Analysis Based On Local Linear Representation And Deep Learning

Medical image segmentation and quantitative measurement of medical image are the crucial issues in medical image analysis,which plays a significant role in the auxiliary diagnosis and assessment of many diseases.Deep learning is a general and effective method for solving many issues of medical image analysis.However,deep learning based model is prone to overfitting when solving medical image analysis problems due to the complicated model which results from the scarce medical image data and the lack of prior knowledge.Local linear representation denotes that a sample on the manifold can be approximated by a linear combination of its several nearest neighbors.Based on manifold assumptions and local linear representation,we can employ some existing prior knowledges(e.g.,local linear mapping,the distribution pattern of the samples from manifold,etc.)to simplify the model.Combining local linear representation and deep learning,we can simplify the model and improve the generalization of the model.Based on local linear representation and deep learning,in this paper,we have made two contributions to the hippocampus segmentation for MR image and the automated quantitative measurement for the spine from MR image:(1)Hippocampus segmentation for MR image based on iterative local linear mapping(ILLM).The hippocampus segmentation problem is divided into a multi-output issue and a threshold segmentation problem.For the multi-output issue,the feature of the MR image patch is used to predict the corresponding distance field(DF)image patch.The absolute value of the voxel in the DF image indicates the closest distance of the voxel to the boundary of the hippocampus.The sign indicates whether the voxel belongs to the hippocampus.After predicting the DF image of the test image,we can obtain the segmentation result by a threshold segmentation approach with the threshold of zero.The deep learning is introduced into the local linear representation model.We propose the semi-supervised deep autoencoder to extract feature from MR image patch.The local structure-preserved manifold regularization(LSPMR)is used to obtained discriminative feature and the anatomical space-constrained dictionary update is exploited to remove outliers of the dictionary,which ensure the locality of feature manifold,the locality of DF manifold,and the consistence of the local structures between the feature manifold and DF manifold.Thus,the hippocampus segmentation accuracy is improved.The experimental dataset is composed of 1.5T T1-weighted MR images from 68 subjects and 3.0T T1-weighted MR images from 67 subjects.Results show that the proposed ILLM achieves accurate hippocampus segmentation with Dice similarity coefficient of 0.8852 and 0.8783 for 1.5T and 3.0T datasets respectively.(2)Automated quantitative measurement of the spine for MR image by cascade amplifier regression network(CARN).The measurements of the height of the lumbar vertebral body and the lumbar intervertebral disc are beneficial to the diagnosis of diseases such as lumbar disc herniation and lumbar disc degeneration.In this paper,CARN achieves accurate and automated measurement of 15 indices of lumbar body height and 15 indices of lumbar disc height by introducing the local linear representation into deep learning model,Cascade amplifier network(CAN)achieves expressive feature embedding and the LSPMR obtains discriminative feature embedding.Moreover,the output of the deep learning model is limited to be close to the output manifold which is captured by local linear representation,which alleviates overfitting.Experimental dataset is composed of Tl-weighted MR images of 215 subjects and T2-weighted MR images of 20 subjects.Results demonstrates that CARN achieves accurate automated quantitative measurement of the spine for MR images with mean absolute error of 1.22mm for 30 indices measurement.