Fig 1.
A flowchart describing an overview of the methodology.
The red highlight marks the automated process output whereas the blue highlight marks the manual process output.
Table 1.
MRI scanning parameters for T2-weighted sagittal scans.
Fig 2.
The process to train the deep learning segmentation model.
The red, green, and blue-highlighted components are used subsequently in the performance analysis step, illustrated in Fig 4.
Fig 3.
An MRI image superimposed with color-coded labels of the six regions illustrating the location of each region in the image.
Fig 4.
The process to analyze the performance of the trained model using the test dataset.
The red, green, and blue-highlighted components are the same highlighted output components in the previous model training step, illustrated in Fig 2.
Fig 5.
An example result of a morphological thinning process, creating a spine line connecting the Sacrum part of the spine to the top part.
Two branches are apparent in this result showing two false end points in addition to the two true end points.
Fig 6.
Illustration of the different components of geometry used to determine the first estimate (magenta line) of the bisecting line segment of an IVD.
Fig 7.
The probability distribution of the pixel intensities belonging to all vertebrae regions in the entire dataset.
Fig 8.
An illustration of the color correlogram of ten unique colors.
Each cell in a matrix, Cij, contains the probability of occurrence of color i and j between two pixels, d distance apart. The yellow-shaded cells are the auto color correlogram, where i is identical to j, whereas the blue-shaded cells represent additional information included in our self-similar color correlogram feature where |i-j|<w and w = 2.
Table 2.
The performance of the best model for each network type.
Table 3.
The values of hyperparameters and training options that were used to train the best model.
Fig 9.
An example of the vertebrae regions labeling result.
Fig 10.
An example of the IVD regions labeling result.
Fig 11.
A visualization of the IVD height measurement stage showing the starting points (magenta), refined points (yellow), final bisecting line segment (yellow dotted line), and the vectors used to shift the points (red and blue arrows).
Fig 12.
Box chart of the relative difference between the predicted and actual IVD heights.
Fig 13.
An example IVD height measurement result that shows the predicted height of three lumbar IVDs.
There is a clear case of trauma at the top plate of the L5 vertebrae causing the algorithm to produce unusually high height prediction for the L4/L5 IVD.
Table 4.
The performance of the Pfirrmann grade classification models.
Table 5.
Range and best hyperparameter values for ensemble binary classification decision tree model.
Fig 14.
The confusion matrix of the Pfirrmann grade classification results.
Grades 1 to 4 were produced using the best Feedforward Fully Connected Neural Network classifier whereas grade 5 used the calculated IVD height.
Fig 15.
The annotation of an L5/S1 IVD with reasonably healthy features where the IVD has a normal height and only limited degenerative characteristics.
Fig 16.
The annotation of an L4/L5 IVD with unhealthy features.
This IVD is annotated with a short IVD height and Pfirrmann grade 5. The blue and red lines in this case do not correspond to the nucleus boundary but rather inner region of the IVD that has relatively brighter pixels than the annulus.