Deep-learning-based biomarker of spinal cartilage endplate health using ultra-short echo time magnetic resonance imaging

Abstract

Background: T2* relaxation times in the spinal cartilage endplate (CEP) measured using ultra-short echo time magnetic resonance imaging (UTE MRI) reflect aspects of biochemical composition that influence the CEP's permeability to nutrients. Deficits in CEP composition measured using T2* biomarkers from UTE MRI are associated with more severe intervertebral disc degeneration in patients with chronic low back pain (cLBP). The goal of this study was to develop an objective, accurate, and efficient deep-learning-based method for calculating biomarkers of CEP health using UTE images.

Methods: Multi-echo UTE MRI of the lumbar spine was acquired from a prospectively enrolled cross-sectional and consecutive cohort of 83 subjects spanning a wide range of ages and cLBP-related conditions. CEPs from the L4-S1 levels were manually segmented on 6,972 UTE images and used to train neural networks utilizing the u-net architecture. CEP segmentations and mean CEP T2* values derived from manually- and model-generated segmentations were compared using Dice scores, sensitivity, specificity, Bland-Altman, and receiver-operator characteristic (ROC) analysis. Signal-to-noise (SNR) and contrast-to-noise (CNR) ratios were calculated and related to model performance.

Results: Compared with manual CEP segmentations, model-generated segmentations achieved sensitives of 0.80-0.91, specificities of 0.99, Dice scores of 0.77-0.85, area under the receiver-operating characteristic curve values of 0.99, and precision-recall (PR) AUC values of 0.56-0.77, depending on spinal level and sagittal image position. Mean CEP T2* values and principal CEP angles derived from the model-predicted segmentations had low bias in an unseen test dataset (T2* bias =0.33±2.37 ms, angle bias =0.36±2.65°). To simulate a hypothetical clinical scenario, the predicted segmentations were used to stratify CEPs into high, medium, and low T2* groups. Group predictions had diagnostic sensitivities of 0.77-0.86 and specificities of 0.86-0.95. Model performance was positively associated with image SNR and CNR.

Conclusions: The trained deep learning models enable accurate, automated CEP segmentations and T2* biomarker computations that are statistically similar to those from manual segmentations. These models address limitations with inefficiency and subjectivity associated with manual methods. Such techniques could be used to elucidate the role of CEP composition in disc degeneration etiology and guide emerging therapies for cLBP.

Keywords: Cartilage endplate (CEP); T2* relaxation time; disc degeneration; low back pain; ultra-short echo time magnetic resonance imaging (UTE MRI).

Figure 1

Flow chart of the individuals included in the study. MRI, magnetic resonance imaging.

Figure 2

UTE image showing lower lumbar CEPs, CEP segmentations, and a transverse T2* map. (A) Mid-sagittal UTE image showing the position of the cropped region, which is centered on the L5 vertebral body. (B) Cropped and up-sampled image with (C) annotated CEP segmentations (red). (D) Transverse T2* map of the inferior L4-L5 CEP. The central CEP region (shown) was used to calculate the mean ± SD T2* relaxation time. UTE, ultra-short echo time; CEP, cartilage endplate; SD, standard deviation.

Figure 3

CEP segmentations (red) were offset into the adjacent disc (orange) and vertebral body (blue) to quantify contrast between the CEP and adjacent tissues. CEP, cartilage endplate.

Figure 4

The trained neural networks produced CEP segmentations and resulting CEP T2* maps that were similar to those generated using manual methods. (A) Representative manually- and (B) model-generated CEP segmentations (red) in a subject from the unseen test dataset. (C) Segmentation overlays showing true positives (+), false positives, and false negatives (−). (D) The transverse T2* maps generated from these segmentations yielded similar estimates of T2* values (mean ± SD) in the central CEP. CEP, cartilage endplate; SD, standard deviation.

Figure 5

Neural network performance as a function of sagittal image position. Mean (A) Dice score and (B) sensitivity were highest mid-sagittally and lowest laterally at all levels (error bars indicate SD). The vertical grey lines annotate the mid-sagittal 50% region, corresponding to the region used to compute the CEP T2* biomarkers and the performance results reported in Table 2 lower. SD, standard deviation; CEP, cartilage endplate.

Figure 6

Level-specific probability cutoff values for binarizing CEP segmentations from predicted probability distributions (likelihood that each pixel was part of the CEP class) were chosen to maximize the 3D Dice score. CEP, cartilage endplate.

Figure 7

Bland-Altman plots of model performance. Compared with manually-generated segmentations, neural-network-generated segmentations provided unbiased estimates of (A) mean CEP T2* values and (B) principal CEP orientation in the unseen test dataset. The majority (60/64, 94%) of model-generated CEP T2* values and angles were within ±4.0 ms and ±5°, respectively, of manually-generated values. CEP, cartilage endplate; SD, standard deviation.