Introduction

Modic changes are variations in magnetic resonance imaging (MRI) signal intensity describing bone marrow lesions adjacent to damaged vertebral body endplates.¹ Bone marrow edemas or inflammation appear distinctly hypointense on T₁-weighted images and hyperintense on T₂-weighted images (Modic type 1). Meanwhile, bone marrow conversion is hyperintense on T₁ MRI and iso- to hyperintense in T₂ images (Modic type 2). And lastly, sclerotic bone appears hypointense in both images (Modic type 3). While Modic changes have been positively associated with lower back pain,^2,3 it is difficult to assess these markers objectively due to inter-operator and inter-scanner variability.⁴ In part, this is due to the qualitative nature of the grading scheme. This study presents an automatic, deep learning approach for detecting the presence of Modic changes in sagittal T₁-weighted and T₂-weighted lumbar spine MRI.

Methods

This Modic mapping scheme consists of three stages: (1) segmentation and localization of the vertebral bodies, (2) binary detection and segmentation of Modic changes, and (3) voxel-wise classification to determine Modic type (Fig. 1). This retrospective study used 75 clinical lumbar spine MRI studies, with and without symptoms of lower back pain. To serve as ground truth for the deep learning component, vertebral bodies with visible Modic changes were segmented for these changes (Type 1, 2 and 3) by two board-certified radiologists. In each MR exam, T₁- and T₂-weighted images were aligned using 3D rigid registration with mutual information loss. A pretrained, in-house segmentation model localized and extracted vertebral bodies with bounding boxes of size 100 × 100. Modic detection was achieved using a second segmentation model which utilized these bounded vertebral bodies and the radiologist-annotated Modic changes. Prior to training the detection model, T₁-weighted and T₂-weighted images of the 1872 sagittal slices of vertebral bodies were converted to z-score maps to normalize for variation in signal intensity and then split into training, validation, and test sets (65/20/15 ratio). This Modic segmentation model was trained using a 2D V-Net⁵ with a combined cross-entropy and Dice loss, learning rate of 1e-5, Adam optimizer, and dropout rate of 0.5. Post-training, the training set was used for cluster analysis, grouping the three types of Modic changes by T₁ z-score and T₂ z-score within the radiologist-annotated segmentations. Finally, Modic maps were generated, where each voxel was classified by the minimal Euclidean distance of T₁ and T₂ z-scores from the cluster centroids. For evaluation, a binary label was assigned to each endplate based upon the presence of voxels of each class. To assess if noise in signal intensity negatively impacted performance, a receiver operating characteristic (ROC) curve was generated using the validation set, with percentage of voxels in each class of the predicted lesions as the varying discrimination threshold. The optimal threshold to maximize $$$True Positive Rate-(1-False Positive Rate)$$$ in the ROC curve was implemented in the final inference of the test set.

Results

The Modic detection model successfully identified the presence or absence of changes in 85.7% of the unseen test set, with a sensitivity of 0.72, specificity of 0.91, and Cohen’s kappa coefficient of 0.63. Voxel-wise classification was derived from clusters centered at [0.23 (±0.73), 1.20 (±1.16)], [1.04 (±1.00), 0.37 (±0.85)], and [-0.53 (±0.41), -0.52 (±0.85)] (Fig. 2). In the assessment of the ROC curve to balance true positive and false positive rates, the optimum threshold was determined to be zero, where the presence of at least one voxel in a class resulted in categorizing the endplate with that class label (Fig. 3). Labeling of each endplate using this rule-based classification system resulted in sensitivities of [0.54, 0.68, and 0.67] and specificities of [0.86, 0.87, and 0.81] for Modic types 1, 2, and 3, respectively (Fig. 4).

Discussion

This study used deep learning-based models to automatically localize and map Modic changes in vertebral bodies. Our results demonstrate substantial agreement of the detection model with radiologist-annotated grading to recognize damaged endplates. By leveraging vertebral body segmentation to mask neighboring spinal structures and signal intensity normalization with z-scoring within the vertebral bodies, the detection model was invariant to external factors in pathology or image acquisition noise. The quantitative nature of this voxel-wise mapping method allows further sensitivity to granular differences and characterizations of image abnormalities. For example, bone marrow pathologies, which have been associated with lower back pain,^6,7 are heterogeneous and often transitional.⁸ Our mapping technique is able to visualize such progression and allow for assessment of these local features (Fig. 5). Furthermore, this model was developed using clinical MRI exams acquired between a period of 8 years, suggesting robustness and generalizability to real-world environments. The limitations of this study are the small and imbalanced sample set with relatively poor representation of Modic type 3. Future stages of this study will explore the implementation of this method to assist radiologist grading.

Conclusion

In this work, we present a novel deep learning-based approach to localize and segment Modic changes, with results that demonstrate high agreement with radiologist grading. The introduction of this fully automatic, quantitative mapping technique may increase inter-operator reliability and ultimately improve robustness in understanding the associations of Modic changes with lower back pain and spinal degeneration.

Acknowledgements

This work is supported by the National Institute of Health and National Institute of Arthritis and Musculoskeletal and Skin Diseases (Project #: UH2AR076724).