Purpose or Case Report: Scoliosis is a deformity of the spine that affects 2-3% of the population. Currently, the reference standard for assessing scoliosis is the manual assignment of Cobb angles at the site of the curvature center using X-ray images. This process is time consuming and unreliable as it is affected by inter- and intra-observer variability. To overcome these inaccuracies, machine learning (ML) methods can be used to automate the Cobb angle measurement process.
Methods & Materials: We propose to address the Cobb angle measurement task using YOLACT, an instance segmentation model. YOLACT is composed of a ResNet backbone that produces a feature pyramid based on the input image. The information inside the feature pyramid is then passed into two parallel pathways. One pathway is the Protonet that performs the segmentation task. The second pathway finds the mask coefficients of the image to produce image masks. Finally, the algorithm linearly combines the two outputs and uses sigmoid nonlinearity to produce the final mask.

After YOLACT finds the segmentation of each vertebra in the X-ray image, using image processing, we extract segmentation contours and detect the minimum bounding box of each segmentation. The four corners of each box are our landmarks. Lastly, we developed an algorithm that iterates through all the extracted landmarks to find the maximum angles.
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. Overall, YOLACT produces acceptable segmentations that follow the curve of the spine with no major deviation. The calculated angles produced by the algorithm were evaluated using two methods. The first method is to use Symmetric Mean Absolute Percentage Error (SMAPE) to measure the relative error in percentage. The proposed method achieves a SMAPE of 10.76%, which outperforms the benchmarks. The second method is to calculate the absolute difference. The majority of the predictions had an error less than 5 degrees and 94.59% of predictions had error less than 10 degrees, indicating the reliability of the YOLACT model.
Conclusions: In comparison to previous methods, the YOLACT approach is reliable and consistent concerning SMAPE with low error rate for predictions. It also does not require an excessive amount of computation resources, demonstrating the promise of an ML tool to contribute to increased accuracy of Cobb angle measurements in the future daily practice of pediatric radiologists.