Purpose or Case Report: X-Ray based manual Cobb angle measurement is inherently time-consuming and associated with high inter- and intra-observer variability. The existing automated scoliosis measurement methods suffer from insufficient accuracy. We propose a two-step segmentation-based deep architecture to automate Cobb angle measurement for scoliosis assessment.
Methods & Materials: The proposed pipeline consists of two phases. First, we utilize a novel augmented U-Net architecture to perform precise segmentation of spinal vertebrae where the target shape is rectangular. Literature suggests the average length of the vertebrae is approximately twice their average height, which was confirmed by our experiments. Hence, we enforced the model to treat the horizontal pixels twice as significant compared to vertical pixels by applying non-square kernels on the augmented path. To further refine the output feature map, we incorporated another feature map produced by square kernels. Consequently, a path with square kernels, which simulated conventional U-Net architecture was employed in parallel. Next, vertices of each vertebra were extracted by minimum bounding boxes to approximate landmark coordinates. Additionally, error correction and outlier rejection techniques were applied to smooth the extracted landmark coordinates by learning from the size and location patterns of adjacent vertebrae.
Results: We trained and tested our model on a cohort of adolescent patients (AASCE-MICCAI 2019 challenge) (n=609) with 80%/20% training and test ratio. The proposed architecture accurately detected all spinal vertebrae in the input X-Ray images and generated precise segmentations, which followed the curve of the spine and clear boundaries between adjacent vertebrae were presented. Successful separation of the adjacent vertebrae improved the precision of the landmark coordinates extraction step. With the AASCE-MICCAI dataset, we achieved the lowest Symmetric Mean Absolute Percentage Error (SMAPE) (9.2%) compared to the benchmarks, with more than 92% of the predictions having less than 10 degrees of absolute difference from the corresponding ground truths, and the majority were less than 5 degrees. We tested our trained model on a separate X-Ray dataset of 55 pediatric patients and achieved similar results.
Conclusions: We proposed augmented U-Net for automated scoliosis measurement in pediatric and adolescent patients. The model outperformed state-of-the-art methods on the AASCE-MICCAI dataset and proved to be robust on an external validation.