Introduction

Low back pain (LBP) is very common worldwide. As many as 70–85% of adults have been estimated to suffer from it at some point in their lives¹. For most (ca. 90%) of patients with LBP, no specific cause for pain can be demonstrated. LBP may result from an injury or degenerative process of the lumbar innervated tissues such as facet joints, intervertebral discs (IVDs), ligaments, or muscles. Magnetic resonance imaging (MRI) shows degenerative changes in vertebrae and IVDs of which some have been associated with the severity of LBP². However, these findings are common in the asymptomatic population as well, and thus may not be representative of clinical symptoms^3,4. Therefore, our study aims to use texture analysis and machine learning methods to find representative features that could predict the presence of LBP from T₂-weighted MRI.

Methods

Mid-sagittal slices (N = 200) of T₂-weighted MRI scans of the lumbar spine along with tabular data containing information about LBP frequency and intensity (scale from 0-10) of Northern Finland Birth Cohort 1966 participants were used in this study. Imaging was done at 1.5 T with a fast spin echo sequence (TE = 112.7 ms, TR = 3500 ms, ETL = 27, slice thickness = 4 mm, matrix size = 512x512, FOV = 280x280 mm²). Based on the pain characteristics at the time of imaging, two groups were formed: subjects with clinically relevant pain (frequency ≥ 30 days and intensity ≥ 6), and no pain (others). Separate regions-of-interest (ROIs) containing the lumbar IVDs and vertebrae were segmented from the images manually (Figure 1). A custom MATLAB (R2019b) script was then used to extract the following textural features from the ROIs in various directions (where applicable): histogram features, gradient features, gray-scale co-occurrence matrix, run-length encoding matrix, wavelet features, autoregressive model parameters, and local binary patterns. In total, 603 textural features were extracted.

Sklearn (v. 0.21.2) machine learning library was used in Python (v. 3.7.3) to build a machine learning pipeline to analyze the feature data. Data were split into training (80%) and test (20%) sets making sure they have the same class distribution. A false-positive-rate test was used to reduce dimensionality by keeping only features testing for p-value below 0.05. A logistic regression classifier was then implemented to predict whether the subject LBP or not. A 5-fold stratified cross-validation scheme was used in conjunction with a grid-search to tune the amount of L₂ regularization in the classifier. The best performing parameters were then used for final classification task. This analysis was done separately for IVD features and vertebral features, and for a combination of both.

Results

The feature reduction algorithm reduced the feature space to 19 IVD features and 22 vertebral features. When using only IVD features, a receiver operating characteristic area under curve (ROC-AUC) of 0.84, accuracy of 0.82, and recall of 0.83 were observed in the test set. For vertebral features only, the corresponding metrics were ROC-AUC = 0.64, accuracy = 0.63, and recall = 0.55. When vertebral and IVD features were combined, the corresponding metrics increased to ROC-AUC = 0.86, accuracy = 0.84, and recall = 0.83. The receiver operating characteristic curves are shown in Figure 2. The receiver operating characteristics for the training and test sets were similar, and therefore our model did not overfit to training data.

Discussion

The results heavily suggest that IVD features have more relevance in predicting clinically relevant LBP. Using only vertebral features resulted in inferior classifier performance. Furthermore, combining features from both ROIs only slightly increased performance when compared to using IVD features only. Indeed, a recent genetic study has shown strongly statistically significant genetic correlation between IVD problems and back pain⁵.

While our results show promise in using texture analysis and machine learning in predicting pain from T₂-weighted images, further investigations are required. Specifically, we aim to increase the amount of analyzed data considerably to ensure good generalization, and to investigate deep learning approaches. In addition, using more MRI contrasts such as T₁ or short-TI inversion recovery (fat suppression) and combining imaging phenotypes such as Modic of Pfirrmann scoring with textural features could further improve the results.

Conclusion

This preliminary work shows that texture analysis combined with a logistic regression classifier pipeline may be used to predict pain from T₂-weighted MR images. Thus, there may be a connection between quantitative MRI features and clinical symptoms in LBP.

Acknowledgements

Support from Technology Industries of Finland Centennial Foundation and Jane and Aatos Erkko Foundation is gratefully acknowledged.