L-MRCA: Landmark-Free Multi-Output Regional Curvature Analysis in Adolescent Idiopathic Scoliosis

Adolescent idiopathic scoliosis (AIS) is a lateral spinal deformity that develops during growth, and its severity is graded by the Cobb angle, which is measured by hand from anteroposterior radiographs. Recent deep learning methods have automated this measurement, improving speed and consistency. However, despite strong reported accuracy, current methods still face two obstacles to deployment where specialists are scarce: (1) dependence on clinician-annotated vertebral landmarks, which are costly to produce and unavailable in most health systems, and (2) localization errors that propagate directly into the reported angle. In this work, we address these challenges by removing the intermediate landmark stage entirely. Specifically, we propose L-MRCA (Landmark-Free Multi-Output Regional Curvature Analysis), which regresses the proximal thoracic, main thoracic, and thoracolumbar Cobb angles directly from a single radiograph using a fully fine-tuned convolutional backbone with a three-output regression head, trained under augmentation constrained to preserve curve laterality. On subset 3 of Spinal-AI2024 dataset, L-MRCA reaches a mean absolute error of 3.70 degrees, with 76.2 percent of predictions within 5 degrees of the reference, and explains most of the variance in the main thoracic region, which is the region that most often determines treatment. What L-MRCA establishes is that regional Cobb angles carry enough signal to be read directly from the image, removing the expert-annotation dependency that confines automated severity grading to well-resourced settings.