Poster Presentation 50th International Society for the Study of the Lumbar Spine Annual Meeting 2024

Lumbar spine T2-weighted MRI radiomics successfully classify intervertebral disc degeneration in the Northern Finland Birth Cohort 1966 (#132)

Terence P McSweeney 1 , Narasimharao Kowlagi 1 , Juhani Määttä 1 , Aleksei Tiulpin 1 2 , Jaro Karppinen 1 3 , Simo Saarakkala 1 4
  1. Research Unit of Health Sciences and Technology, University of Oulu, Oulu, Northern Ostrobothnia, Finland
  2. Neurocenter Oulu, University of Oulu, Oulu, Northern Ostrobothnia, Finland
  3. Rehabilitation Services of South Karelia Social and Health Care District, Lappeenranta, South Karelia, Finland
  4. Department of Diagnostic Radiology, University Oulu Hospital, Oulu, Northern Ostrobothnia, Finland

Introduction

Multiple quantitative measures of intervertebral disc (IVD) degeneration from T2-weighted MRI have been investigated from manual or semi-automatic segmentation. Of these, only disc height index (DHI) and delta peak signal intensity (DPSI) are consistently reported as potential imaging biomarkers for IVD degeneration [1].

In this study, we aim to use deep learning (DL) segmentations to extract these measurements and numerous additional imaging features, collectively termed radiomics, from a population cohort. Subsequently, we aim to identify a set of features that best classify IVD degeneration and compare these to the performance of DHI and DPSI.

Methods

We used data from Northern Finland Birth Cohort 1966 members who underwent lumbar spine T2-weighted MRI scans at age 46-47 and had Pfirrmann grade evaluations (n=1407). We used a previously published DL model [2] to segment the lumbar spine IVDs from sagittal slices. We tested the quality of these segmentations against a set of 300 IVDs manually annotated by an experienced musculoskeletal researcher (JM). We extracted a set of 465 radiomic features using PyRadiomics [3]. We also calculated DHI [4] and DPSI [5].

Data were split into development (80%) and test (20%) sets stratified by participant, Pfirrmann grade, and pain. Intraclass correlation coefficients values for features from DL versus manual segmentations greater than 0.95 were kept. Principal component analysis (PCA) and a support vector machine were used to train a Pfirrmann grade classification model. Separate models were trained using DHI and DPSI and radiomic features individually as well as in combination. Finally, PCA contributions were examined to identify the most relevant radiomic features.

Results

Out of 300 manually annotated IVDs, the DL model misidentified the level in 17 cases. Excluding these, average agreement between the experienced musculoskeletal researcher (JM) and the DL segmentations was as follows: Dice coefficient 0.90 (0.87-0.94), Jaccard index 0.82 (0.77-0.88) and 95% Hausdorff of 2.12 mm (0.60 mm-3.65mm). ICC values and 95% confidence intervals for DHI and DPSI were 0.83 (0.52-0.92) and 0.63 (0.17-0.81), respectively. Out of 465 radiomic features, 79 had an ICC greater than 0.95.

In the test set, the model using all features achieved balanced accuracy of 77.86% (74.15%-81.11%) and Cohen’s Kappa of 0.70 (0.67-0.73), compared to 65.19% (61.15%-68.88%) and 0.55 (0.51-0.58), respectively, for DHI and DPSI alone. Gray level dependency and run length matrix features, were the primary contributors to the first dimension of the PCA (Figure 1).

65521544d34d8-Screenshot+2023-11-13+at+12.20.10.png

Figure 1. Scatter plot of the first two dimensions of the PCA. The top 5 contributing features are shown along with individual IVDs and their Pfirrmann classification. Abbreviations: LoG = Laplacian of Gaussian; gldm = gray level dependence matrix; glrlm = gray level run length matrix.

Discussion

This work shows the utility of radiomics from DL IVD segmentation in degeneration classification, with features of IVD homogeneity having greater importance. Interestingly, radiomic features afforded a significant advantage over commonly used quantitative features, possibly due to higher reliability of radiomic features versus DHI and DPSI. We believe these radiomic features warrant further investigation as potential imaging biomarkers of IVD degeneration.

  1. Pereira CL, Grad S, Gonçalves RM. Biomarkers for intervertebral disc and associated back pain: From diagnosis to disease prognosis and personalized treatment. JOR Spine. 2023;e1280.
  2. Kowlagi N, Nguyen HH, McSweeney T, Saarakkala S, Määttä J, Karppinen J, Tiulpin A. A Stronger Baseline For Automatic Pfirrmann Grading Of Lumbar Spine Mri Using Deep Learning. 2023 IEEE 20th Int Symp Biomed Imaging (ISBI). 2023;00:1–5.
  3. Griethuysen JJM van, Fedorov A, Parmar C, Hosny A, Aucoin N, Narayan V, Beets-Tan RGH, Fillion-Robin JC, Pieper S, Aerts HJWL. Computational Radiomics System to Decode the Radiographic Phenotype. Cancer Res. 2017;77(21):e104–e107.
  4. Videman T, Battié MC, Gibbons LE, Gill K. Aging changes in lumbar discs and vertebrae and their interaction: a 15-year follow-up study. Spine J. 2014;14(3):469–478.
  5. Waldenberg C, Hebelka H, Brisby H, Lagerstrand KM. MRI histogram analysis enables objective and continuous classification of intervertebral disc degeneration. Eur Spine J. 2018;27(5):1042–1048.