Automatic Vertebral Body Segmentation Based on Deep Learning of Dixon Images for Bone Marrow Fat Fraction Quantification

Abstract

Background: Bone marrow fat (BMF) fraction quantification in vertebral bodies is used as a novel imaging biomarker to assess and characterize chronic lower back pain. However, manual segmentation of vertebral bodies is time consuming and laborious. Purpose: (1) Develop a deep learning pipeline for segmentation of vertebral bodies using quantitative water-fat MRI. (2) Compare BMF measurements between manual and automatic segmentation methods to assess performance. Materials and Methods: In this retrospective study, MR images using a 3D spoiled gradient-recalled echo (SPGR) sequence with Iterative Decomposition of water and fat with Echo Asymmetry and Least-squares estimation (IDEAL) reconstruction algorithm were obtained in 57 subjects (28 women, 29 men, mean age, 47.2 ± 12.6 years). An artificial network was trained for 100 epochs on a total of 165 lumbar vertebrae manually segmented from 31 subjects. Performance was assessed by analyzing the receiver operating characteristic curve, precision-recall, F1 scores, specificity, sensitivity, and similarity metrics. Bland-Altman analysis was used to assess performance of BMF fraction quantification using the predicted segmentations. Results: The deep learning segmentation method achieved an AUC of 0.92 (CI 95%: 0.9186, 0.9195) on a testing dataset (n = 24 subjects) on classification of pixels as vertebrae. A sensitivity of 0.99 and specificity of 0.80 were achieved for a testing dataset, and a mean Dice similarity coefficient of 0.849 ± 0.091. Comparing manual and automatic segmentations on fat fraction maps of lumbar vertebrae (n = 124 vertebral bodies) using Bland-Altman analysis resulted in a bias of only -0.605% (CI 95% = -0.847 to -0.363%) and agreement limits of -3.275% and +2.065%. Automatic segmentation was also feasible in 16 ± 1 s. Conclusion: Our results have demonstrated the feasibility of automated segmentation of vertebral bodies using deep learning models on water-fat MR (Dixon) images to define vertebral regions of interest with high specificity. These regions of interest can then be used to quantify BMF with comparable results as manual segmentation, providing a framework for completely automated investigation of vertebral changes in CLBP.

Keywords: biomarkers; bone marrow fat; deep learning; magnetic resonance imaging; segmentation; spine imaging.

Figure 1

Automatic vertebrae segmentation and fat quantification pipeline. (A) All IDEAL images (water, fat, fat fraction, and ${\text{R}}_2^{*}$) are fed into a U-Net (13) as multichannel inputs, resulting in the predicted segmentation map. Each ROI corresponding to lumbar vertebrae was analyzed on fat fraction maps to yield mean BMF values. (B) DICOM masks were made from the predicted segmentation map. ROIs were identified through the MaskToMir function in in-house software made in IDL (IDL, Research Systems, Broomfield, CO). These automatically identified ROIs were then overlaid on the fat fraction maps derived from the water-fat IDEAL image series. For each lumbar vertebral body ROI, the mean fat fraction value was obtained for each slice, and the final mean BMF was averaged over all slices with lumbar vertebral bodies present. The mean BMFs for each lumbar vertebral body as defined by the automatically segmented ROIs were compared with the manually segmented ROIs through Bland-Altman analysis.

Figure 2

Description of datasets evaluated in this study. Both training sets consisted of the same 31 subjects. Set 1A was used to train the U-Net. Automatic segmentations from Set 1A were compared with manual segmentations done by Rater A and also used during the U-Net training process. The U-Net with finalized weights after training was then used to automatically segment images from the same image set 1 and compared with manual segmentations done by Rater B, denoted as Set 1B. Set 2A consisted of 11 subjects, with 7 slices each with identified vertebra from Rater A's manual segmentation, while Set 2B&3B consisted of 26 subjects, with 5 slices each with identified vertebra from Rater B's manual segmentation. Additional demographic information can be found in Table 1.

Figure 3

Bland-Altman plots of mean bone marrow fat fraction (BMF) percentages (%) as determined by manual segmentation compared to automatic segmentation for each lumbar vertebral body (L1-L5). The biases between BMFs collected by the automated (NN) and manual segmentations for both test sets were less than 10% of the mean value. (A) Comparison of mean BMF values for manual segmentations performed by annotator A and the predicted segmentation by the deep learning model on Set 2A (n = 53 vertebrae). The bias was +0.382% with limits of agreement of −1.850% and +2.614%. (B) Comparison of mean BMF values for manual segmentations performed by annotator B and the predicted segmentation by the deep learning model on Set 2B&3B (n = 124 vertebrae). The bias was−0.605% with limits of agreement of −3.275% and +2.065%.