Fig 1.
A T1-weighted (left) and a T2-weighted (right) traverse MRI images of the L3/L4 Intervertebral Disc of a patient are shown.
One marked difference in the two images is the cerebrospinal fluid (CSF) in the spinal canal that appears black on the T1-weighted image but as a brighter region on the T2-weighted image because of its low fat contents.
Fig 2.
A mid-sagittal view of a lumbar spine MRI showing the intersection lines between the sagittal plane and the traverse planes that are shown in Fig 3.
The lines marked in red are the intersection lines of traverse planes that cut closest to the half-height of an IVD.
Fig 3.
An example of nine traverse images of a lumbar spine.
Image 2, 5, and 8 are from the planes that cut closest to the half-height of L3/L4, L4/L5, and L5/S1 IVD, respectively.
Fig 4.
A flowchart describing the traditional Transfer Learning approach of using Deep Convolutional Neural Network for medical image classification, where a) depicts the training process and b) depicts the inference step.
Table 1.
The range of acquisition parameter values used during traverse MRI scans.
Fig 5.
Two example cases where the image registration process succeeded (left column) and failed (right column).
The top row shows the T1-weighted images, the middle row shows T2-weighted images, and the bottom row shows the resulting composite images after image registration.
Table 2.
Dataset sizes (number of images).
Fig 6.
A flowchart describing the methodology used where a) depicts the feature extraction step, b) depicts the DR and FS modeling step, c) depicts the ML and FC training step and d) depicts the inference/classification step.
Table 3.
The list of the DCNNs used and the summary of their architecture.
Fig 7.
A tree diagram depicting the method combination used in this study.
Table 4.
The description of the hyperparameter optimization options and range of values for each ML learner.
Table 5.
The description of the hyperparameter optimization options and values for each FC learning optimization algorithm.
Table 6.
The feature length of each DCNN model and DR/FS method combination.
Table 7.
The best ML learner for each combination method and its average classification performance (using Accuracy metric).
Table 8.
The best ML learner for each combination method and its average classification performance (using Precision metric).
Table 9.
The best ML learner for each combination method and its average classification performance (using Recall metric).
Table 10.
The best ML learner for each combination method and its average classification performance (using F1-Score metric).
Table 11.
The best FC learning optimizer for each combination method and its average classification performance (using Accuracy metric).
Table 12.
The best FC learning optimizer for each combination method and its average classification performance (using Precision metric).
Table 13.
The best FC learning optimizer for each combination method and its average classification performance (using Recall metric).
Table 14.
The best FC learning optimizer for each combination method and its average classification performance (using F1-Score metric).
Table 15.
Summary of the classification performance of ML learners using features from 11 DCNNs.
Table 16.
Summary of the classification performance of FC neural networks using features from 11 DCNNs.
Table 17.
Summary of classification performance of ML methods for each DCNN.
Fig 8.
The Accuracy of DR/FS/FL-ML methods using DenseNet201 features.
The ML learners used can be found in the second row of Table 7.
Fig 9.
The Precision of DR/FS/FL-ML methods using DenseNet201 features.
The ML learners used can be found in the second row of Table 8.
Fig 10.
The Recall of DR/FS/FL-ML methods using DenseNet201 features.
The ML learners used can be found in the second row of Table 9.
Fig 11.
The F1-Score of DR/FS/FL-ML methods using DenseNet201 features.
The ML learners used can be found in the second row of Table 10.
Fig 12.
Per-class classification performance using Fine Gaussian SVM classifier on full-length DenseNet201 features (Precision).
Fig 13.
Per-class classification performance using Fine Gaussian SVM classifier on full-length DenseNet201 features (Recall).
Fig 14.
Per-class classification performance using Fine Gaussian SVM classifier on full-length DenseNet201 features (F1-Score).