90 - Are there differences in paraspinal muscle cross-sectional areas in patients with myopathic or degenerative scoliosis? A deep learning approach

Introduction
The current method of differentiating the origin of scoliosis as myopathic or degenerative is through muscle biopsy and electromyography, which can be time-consuming. Identification of the origin of scoliosis is important for surgical planning. This is a feasibility study that uses a deep learning model to calculate the cross-sectional areas (CSA) of paraspinal muscles to identify differences in patients with and without scoliosis.

Methods
This was a single institution, retrospective study of patients with biopsy-confirmed myopathic scoliosis, with degenerative scoliosis, or without scoliosis. Patients were matched based on baseline characteristics. A convolutional neural network segmentation model trained on sagittal and axial T1-weighted MRIs was used to calculate the CSA of the paraspinous muscles including multifidus, erector spinae, quadratus, and psoas. The paraspinal muscles were also graded manually using the simplified Goutallier classification based on degree of fat infiltration: grade 0 for 0-10%, grade 1 for 10-50%, and grade 2 for >50%. Statistical analysis was conducted to correlate the CSAs to the manual quantitative muscle grades.

Results
Of the 37 patients included in the study, 11 (29.7%) had myopathic scoliosis, 14 (37.8%) had degenerative scoliosis, and 12 (32.4%) did not have scoliosis. The average age was 67.8 years and 59.5% were female. In those with myopathic scoliosis, the model computed an average CSA of 627.4±499.2cm2 for grade 2, 574.1±463.9cm2 for grade 1, and 548.8±357.8cm2 for grade 0. In those with degenerative scoliosis, the average CSA was 685.3±455.1cm2 for grade 2, 596.8±462.2cm2 for grade 1, and 515.5±263.8cm2 for grade 0. In those without scoliosis, the average CSA was 550.7±647.9cm2 for grade 2, 938.3±1111.6cm2 for grade 1, and 442.6±424.9cm2 for grade 0. The difference between grade 2 and 0 CSAs was significant (p < 0.05) but there were no differences in different gradings amongst the three cohorts.

Conclusion
The deep learning model can be used to reliably distinguish between the different muscle gradings based on the extent of fat infiltration regardless of spine condition. The model can be further refined to differentiate amongst the different types of muscles and levels to ultimately identify differences in patients with and without scoliosis of different origins.