Construction Of Deep Models For Quantitative Indices Estimation From Spine Image

Spine image quantification is to measure some quantitative indices from spine images.For example,axial spine image quantification measures four indices that are associated with the dural sac or intervertebral disc,including dural sac mid-sagittal diameter,dural sac cross-sectional area,intervertebral disc mid-sagittal diameter,intervertebral disc cross-sectional area.Automatic estimation of axial spine indices is clinically desired for various spine computer aided procedures,such as disease diagnosis,therapeutic evaluation,pathophysiological studies,risk assessment,and biomechanical modeling.Currently,the spine indices are manually measured by physicians,which is time-consuming and laborious.Even worse,the tedious manual procedure might result in inaccurate measurement.To deal with this problem,we proposed two deep models for firstly automatic estimation of axial spine image.(1)Multiple Axial Spine Indices Estimation via Dense Enhancing Network(DE-Net).Inspired by the success of deep learning for regression problems and the dense convolutional network(DenseNet)for image classification,we propose a dense enhancing network(DE-Net)which uses the dense enhancing blocks(DEBs)as its main body,where a feature enhancing layer is added to each of the bypass in a dense block.The DEB is designed to enhance feature response of the intervertebral disc and the dural sac areas.In addition,the cross-space distance-preserving regularization(CSDPR),which enforces consistent inter-sample distances between the output and the ground-truth spaces,is proposed to regularize the loss function of the DE-Net.(2)Axial Spine Indices Estimation via Object-Specific Bi-Path Network(OSBP-Net).Based on the observation of the size-feature of target organs,observation of image grey value,exploration of spatial relationship between target organs and exploration of relation between estimated indices,we constructed the OSBP-Net.Considering target organs include intervertebral disc and dural sac,OSBP-Net is constructed as a bi-path network.We proposed shallow feature extraction layer(SFE)and deep feature extraction sub-network(DFE)for discriminative feature from specific target organ(intervertebral disc or dural sac).The two SFEs use different convolution strides based on the consideration that the two target organs have different anatomical sizes.The DFEs use average pooling for downsampling based on the observation that the target organs have lower intensity than background.In addition,an inter-path dissimilarity constraint(IPDC)is proposed and applied to the output of the SFEs based on the consideration that the activated regions in the features maps of two paths should be different theoretically,which eliminates background and object disturbances by excluding the overlapped regions of features maps from two paths.Finally,an inter-index correlation regularization(IICR)is proposed and applied to the last fully connected layer of the DFEs based on the observation that the diameter and area of the same object express an approximately linear relation,which exploits the correlations between indices by modeling relation between indices.The design of OSBP-Net and its loss function fully considers the intrinsic properties of the axial spine MRI image and the quantitative indices,and therefore results in high prediction accuracy.To train and validate the proposed method,we collected 895 axial spine magnetic resonance imaging(MRI)images from 143 subjects.The results of the proposed networks were compared with that of several conventional machine learning methods and deep learning methods.It was demonstrated that the proposed DE-Net and OSBP-Net acquire the smallest error among all competing methods,indicating that our method has great potential for spine computer aided procedures.Specifically,the prediction accuracy of the DE-Net for DSMD,DSCA,IDMD and IDC A is 13.94%,13.17%,18.92%and 1.25%higher than that of the best comparison method.In addition,the prediction average absolute error of the DE-Net of the four indicators are 1.04 mm,2.36 mm2,1.47 mm and 3.54 mm2.The prediction errors of the OSBP-Net for the four indicators are smaller than the DE-Net,which are 0.87 mm,2.18 mm2,1.27 mm and 2.98 mm2,respectively.In conclusion,this thesis aims at using deep learning technique to automatically predict indices from axial spine image.Contributions and innovations of this study can be summarized as follows:1)Axial spine image quantification is of great importance and there has been no literature reporting deep learning methods for such task.This thesis presents the first study investigating the feasibility of using deep learning for axial spine image quatification.2)Based on in-depth analyses of the spine MRI image and the corresponding quantitative indices,we proposed two models,the DE-Net and the OSBP-Net,and specifically designed regularizations for network training to acquire high prediction accuracy.3)To train and validate the proposed deep models,we collected a dataset including 895 clinical axial spine MRI images,conducted extensive experiments based on the dataset and made numerous comparisons between the proposed models and several other models.We demonstrated that the proposed models could obtain very promising results.