Deep learning for automated, interpretable classification of lumbar spinal stenosis and facet arthropathy from axial MRI

Abstract

Objectives: To evaluate a deep learning model for automated and interpretable classification of central canal stenosis, neural foraminal stenosis, and facet arthropathy from lumbar spine MRI.

Methods: T2-weighted axial MRI studies of the lumbar spine acquired between 2008 and 2019 were retrospectively selected (n = 200) and graded for central canal stenosis, neural foraminal stenosis, and facet arthropathy. Studies were partitioned into patient-level train (n = 150), validation (n = 20), and test (n = 30) splits. V-Net models were first trained to segment the dural sac and the intervertebral disk, and localize facet and foramen using geometric rules. Subsequently, Big Transfer (BiT) models were trained for downstream classification tasks. An interpretable model for central canal stenosis was also trained using a decision tree classifier. Evaluation metrics included linearly weighted Cohen's kappa score for multi-grade classification and area under the receiver operator characteristic curve (AUROC) for binarized classification.

Results: Segmentation of the dural sac and intervertebral disk achieved Dice scores of 0.93 and 0.94. Localization of foramen and facet achieved intersection over union of 0.72 and 0.83. Multi-class grading of central canal stenosis achieved a kappa score of 0.54. The interpretable decision tree classifier had a kappa score of 0.80. Pairwise agreement between readers (R1, R2), (R1, R3), and (R2, R3) was 0.86, 0.80, and 0.74. Binary classification of neural foraminal stenosis and facet arthropathy achieved AUROCs of 0.92 and 0.93.

Conclusion: Deep learning systems can be performant as well as interpretable for automated evaluation of lumbar spine MRI including classification of central canal stenosis, neural foraminal stenosis, and facet arthropathy.

Key points: • Interpretable deep-learning systems can be developed for the evaluation of clinical lumbar spine MRI. Multi-grade classification of central canal stenosis with a kappa of 0.80 was comparable to inter-reader agreement scores (0.74, 0.80, 0.86). Binary classification of neural foraminal stenosis and facet arthropathy achieved favorable and accurate AUROCs of 0.92 and 0.93, respectively. • While existing deep-learning systems are opaque, leading to clinical deployment challenges, the proposed system is accurate as well as interpretable, providing valuable information to a radiologist in clinical practice.

Keywords: Arthropathy; Deep learning; MRI, Stenosis.

Fig. 1

Imaging examples of T2-weighted axial MRI graded as normal, mild, moderate, or severe for central canal stenosis (A), normal/mild or moderate/severe for neural foraminal stenosis (B), and normal/mild or moderate/severe for lumbar facet arthropathy (C)

Fig. 2

Overview of the deep learning pipeline. T2-weighted axial slices are passed into V-Net segmentation models to obtain masks for the intervertebral disc and dural sac. Geometric rules based on the disc and dural sac are used to localize bounding boxes around foramen and facet. Each localized region is passed into its corresponding classifier: Big Transfer (BiT) convolutional neural network (CNN) for classification of lumbar spinal stenosis, foraminal stenosis, and facet arthropathy. Interpretable classification (decision tree) of lumbar spinal stenosis relies on additional quantitative metrics extracted from the disc and dural sac segmentations

Fig. 3

Results of the V-Net segmentation model on a T2-weighted MR axial slice at L2/L3 from the test set. Model predictions (red) and ground-truth annotations (green) show significant overlap with Dice scores of 0.91 and 0.96 for the intervertebral disc and the dural sac, respectively. Also illustrated are the generated bounding boxes (red) compared to ground truth (green) for the right facet and left foramen

Fig. 4

Receiver operator characteristic curve (ROC) of the BiT convolutional neural network (magenta) and the interpretable decision tree classifier (green) for binary classification of lumbar spinal stenosis with their respective area under the ROC (AUROC) values reported in the legend. The difference in AUROC was statistically significant (p < 0.01) based on DeLong’s paired test for AUROC

Fig. 5

Receiver operator characteristic curve (ROC) of the BiT convolutional neural network for binary classification of foraminal stenosis (A) and facet arthropathy (B)

Fig. 6

Visualization of the predictions of the landmark coordinate regression model on the mid-sagittal slice of the T2 sequence. Green (ground-truth landmark) and red (model-predicted landmark) had a mean absolute error of 1.14 mm in this example. Since the model is constrained to generate exactly 5 landmarks, they correspond to L1/L2, L2/L3, L3/L4, L4/L5, and L5/S1