Introduction: Automated MRI systems such as SpineNet¹ enable rapid re-annotation of cohort intervertebral disc traits onto the same objective scale¹, independent of the original MRI annotation systems², and hence allow direct comparisons of degenerative disc gradings between cohorts. Here, we aimed to increase the predictive capability of SpineNet, a Convolutional Neural Network (CNN) model, by training it to examine differences in morphology in images of the same individual taken 10 years apart, and in differences in morphology between individuals. To assess the sensitivity of the system we examined differences between vertebral bodies (volumes) of middle-aged subjects, assuming they were less likely to change morphologically than the intervertebral discs. 

Methods: The MRI dataset comprised 920 consented subjects (6327discs) from Twins UK (https://twinsuk.ac.uk/). 423 of these subjects have follow-up scans 10-12 years after baseline imaging; these scans were used to examine intra-individual spinal differences.

A model was trained solely on pairs of 3D vertebral body (VB) volumes at each level of the lumbar area (T12 to L5); “same” pairs are pairs sampled from the same individual at different timepoints while “different” pairs are pairs sampled from different individuals ignoring timepoint information. Monozygotic/dizygotic twin pairs were omitted during training. The model was trained to produce a measure of distance (i.e. difference) between the morphology of two VBs, such that the lower the distance, the more similar were the two vertebrae.  While distances were calculated per VB pair, here we show the distances summed to give a measure of the distance between two spines i.e. subjects.

Results: This model was sufficiently sensitive to detect differences in morphology of paired vertebrae. The distances for all spine pairings in our dataset, is shown in Figure 1.  MR images of the same individual taken 10-12 years apart, had the smallest distance between them, i.e. were most similar. For scans taken at the same time point, as expected, the distance between monozygotic twins, was smaller than between dizygotic twins, with the distance between unrelated individuals, noticeably larger than that between any of their twin pairs.

Conclusions: This study shows a deep learning model can be trained to detect and quantify morphological differences of spinal features such as vertebral bodies in terms of image ‘distance’.  Here we showed it can be used to detect ‘distances’ between individuals, and ‘distances’ between the same individual, imaged at different times.  This latter capability would be of particular benefit in longitudinal imaging studies, as it would enable quantitative assessment of morphological changes in the spine.

References:

1.Jamaludin et al; 2017. doi.org/10.1007/s00586-017-4956-3

2. Sambrook et al,1999, Arthritis Rheum 1999;42:366–372