Quantitative framework for the diagnostic, prognostic, and therapeutic evaluation of spinal cord diseases
A method of generating a quantitative characterization of injury presence and status of spinal cord tissue using an adaptive CNN system for use in diagnostic assessment, surgical planning, and therapeutic strategy comprises preprocessing for artifact correction of diffusion based, spinal cord MM data, training an adaptive CNN system with healthy and abnormal (injured/pathologic) spinal cord images obtained by imaging a population of healthy, typically developed spinal cord subjects and subjects with spinal cord injury, evaluating a novel, diffusion-based MM image for injury biomarkers using the adaptive CNN system, generating a three-dimensional predictive axonal damage map for quantitative characterization and visualization of the novel, diffusion-based MM image, and transmitting the sets of healthy and injured spinal cord images back to a central database for continued improvement of the adaptive CNN system training. A system for defining a predictive spinal axonal damage map is also described.
1 . A system for defining a predictive spinal axonal damage map comprising:
a processor; a memory device comprising instructions, which when executed by the processor, perform steps comprising:
accessing a patient derived data comprising an MRI image;
using the patient derived data as an input to an adaptive CNN system, the adaptive CNN system defined to create a set of normative diffusion and texture values of said derived data;
comparing the patient derived data to the normative diffusion and texture values; and
generating a three-dimensional predictive spinal axonal damage map for quantitative characterization and visualization of the MRI image.
2 . The system of claim 1 , wherein said MRI image is preprocessed for artifact correction of diffusion based spinal cord MRI data.
3 . The system of claim 1 , wherein the patient derived data is of a healthy spinal MRI image.
4 . The system of claim 1 , wherein the patient derived data is an MRI image of a damaged spinal cord.
5 . The system of claim 1 , the steps further comprising:
utilizing a preprocessing scheme for artifact correction of diffusion based, spinal cord MRI data;
training an adaptive CNN system with healthy and abnormal spinal cord images obtained by imaging a population of healthy, typically developed spinal cord subjects and subjects with spinal cord injury;
evaluating a novel, diffusion-based MRI image for injury biomarkers using the adaptive CNN system;
generating a three-dimensional predictive spinal axonal damage map for quantitative characterization and visualization of the novel, diffusion-based MRI image; and
transmitting the sets of healthy and injured spinal cord images back to a central database for continued improvement of the adaptive CNN system training.
6 . The system of claim 5 , the steps further comprising:
capturing and storing diffusion-sensitized MRI data, the data including at least one of: Diffusion Tensor Imaging (DTI) data, Diffusion Kurtosis Imaging (DKI) data, Neurite Orientation Dispersion and Density Imaging (NODDI) data, and Other higher order diffusion sampling schema;
correcting artifacts of the diffusion-based MRI data which can but does not explicitly mandate the inclusion of noise reduction, eddy current correction, segmentation of the spinal cord, ghost removal, coregistration and motion correction, and automated image quality control;
estimating and calculating the diffusion tensor, kurtosis tensor, and multi-compartment models using an iteratively-reweighted outlier rejection scheme;
analyzing a texture to extract relevant 1 st order, 2 nd order, higher order, as well as geometric image features for texture training of the adaptive CNN system; and
including healthy volunteer and injured spinal cord datasets, processed as detailed above, for continued diagnostic training of the adaptive CNN system.
7 . The system of claim 5 , further comprising: wherein said training data set comprises at least 100 MRI images of healthy spinal cords and at least 100 MRI images of abnormal spinal cords.
8 . The system of claim 5 , wherein the diffusion sensitized data is captured and stored.
9 . The system of claim 5 , wherein the pre-processing comprises a step of noise reduction based on a principal component analysis (PCA) method for signal decomposition and elimination of noise.
10 . The system of claim 5 , wherein the pre-processing comprises a step of eddy current correction based on estimation of the eddy current (EC) distortion field.
11 . The system of claim 5 , wherein the pre-processing comprises a step of segmentation of the cord using histogram thresholding for tissue classification, morphological processing, and region growing techniques.
12 . The system of claim 5 , wherein the pre-processing comprises a step of ghost removal based on texture analysis and automated classification using neuro-adaptive fuzzy logic.
13 . The system of claim 5 , wherein the pre-processing comprises a step of co-registration and realignment based on maximized mutual information and a variable degree of freedom transformation.
14 . The system of claim 5 , wherein the pre-processing comprises an automated image quality control step based on local SNR characteristics, Euclidean metrics, and tissue contrast.
15 . The system of claim 5 , wherein the pre-processing comprises a tensor estimation step unique to the sampling scheme used.
16 . The system of claim 5 , wherein the pre-processing comprises an outlier rejection step based on iteratively reweighting individual diffusion-sensitized images based on residuals during the tensor estimation process.
17 . The system of claim 5 , wherein after completion of a pre-processing, a pattern recognition phase is performed.
18 . The system of claim 17 , wherein the pattern recognition phase comprises histogram textures comprising at least one of image intensity, variance, entropy, skewness, or kurtosis.
19 . The system of claim 17 , wherein the pattern recognition phase comprises 2 nd order textures comprising at least one of co-occurrence matrix contrast, homogeneity, correlation, or energy at 45° intervals.
20 . The system of claim 17 , wherein the pattern recognition phase comprises higher order textures comprising at least one of mean, variance, entropy, or energy of vertical, horizontal, or diagonal wavelet components.
21 . The system of claim 17 , wherein the pattern recognition phase comprises geometric textures comprising at least one of edge, corner, roundness, elasticity, or solidness.
22 . The system of claim 17 , wherein the pattern recognition phase comprises histogram textures, 2 nd order textures, higher order textures, and geometric textures, and statistical testing of the same to determine feature relevance for the adaptive CNN system.
23 . The system of claim 4 , where training an adaptive CNN system using the captured and processed diffusion sensitized MR data for normal and abnormal subjects.
24 . The system of claim 5 , wherein the trained adaptive CNN system classifies imaged tissue from patients as likely to be injured or not likely to be injured based on the aforementioned training using normative diffusion sensitized MR data.
25 . The system of claim 5 , wherein the steps further comprise generating a quantitative, three-dimensional predictive spinal axonal damage map for the spinal cord; and
wherein output from the trained, adaptive CNN system will provide the quantitative information needed to characterize the presence, location, and extent of spinal cord injury on a per subject basis.
26 . The system of claim 5 , the steps further comprising identifying injury to a spinal cord, wherein the predictive spinal axonal damage map will be used by clinicians, surgeons, physical therapist, PM&R staff for reporting, surgical planning, and therapeutic strategy, wherein the map is used to precisely localize candidate sites for gene therapy with greater specificity than is possible through conventional MR imaging as well as the tracking of therapeutic outcomes by examining changes in microstructural diffusion properties over the course of treatment.
27 . The system of claim 5 , wherein the predictive spinal axonal damage map is imported to a neuronavigational system;
wherein the map is imported as a 3-D image to the neuronavigational system; and
wherein the map is overlaid onto conventional, structural MR imaging to provide the clinician with a real time view, location, and extent of tissue damage during surgical procedures.
28 . The system of claim 5 , wherein the diffusion sensitized image features are transmitted and stored on a centralized database for continued improvement of the adaptive CNN system training.