Medical Student David Geffen School of Medicine at UCLA Los Angeles, California, United States
Introduction: Syringomyelia, or syrinx, refers to the pathologic formation of a fluid-filled cyst within the spinal cord. It is known that clinical severity of associated symptoms correlates with cystic volume and surgical intervention commonly aims to shrink syrinx volume, yet there exists no fully automatic method for quantifying syrinx volume. We introduce the first machine learning model to automatically segment syrinx and compute syrinx volume based on axial T2-spinal MRI.
Methods: Twenty-nine T2-weighted axial spinal MRIs were downloaded from our institutions picture archive system and immediately anonymized. A team of expert clinicians then segmented the borders of the syrinx cavities throughout a set of 19 MRI-series, serving as a training set, upon which a two-stage deep-U-net ensemble was trained. Two independent clinician raters (R1 and R2) and the algorithm (A1) then independently segmented syrinxes in a hold-out test set of 10 separate MRI-series. Dice scores, Hausdorff coefficients, average surface distances (ASDs), and percentage error in segmented syrinx volume were then calculated for A1/R1, A1/R2, and R1/R2.
Results: Average dice scores were 0.78±0.05 (±1 SD), 0.74±0.09, and 0.82±0.08 for A1/R1, A1/R2, and R1/R2 respectively. Average Hausdorff coefficients were 10.31±9.31, 13.03±8.57, and 8.09±9.45 for A1/R1, A1/R2, and R1/R2 respectively. Average surface distances were 0.98±0.09, 1.03±0.05, and 0.93±0.07 for A1/R1, A1/R2, and R1/R2 respectively. Average percent syrinx volume errors were 14.31±3.45%, 15.28±4.98%, and 12.87±4.98% for A1/R1, A1/R2, and R1/R2 respectively.
Conclusion : We present the first machine learning model capable of automatically segmenting syrinx on T2-spinal MRI and calculating syrinx volume. This algorithm achieved accuracy comparable to that of expert clinician raters over a small test set. This work serves as a basis for integrating more objective assessment of the temporal evolution of syrinx cavities from both a morphologic and volumetric standpoint. Future research will strive to further improve segmentation accuracy and validate results in a larger dataset.
