Automated non-contact measurement of the spine curvature at the sagittal plane using a deep neural network

Abstract

Background: Non-radiographical techniques have been suggested to measure the spine curvature at the sagittal plane. However, a neural network has not been used to measure the curvature.

Methods: A single video camera captured images of a standing posture at the sagittal plane from twenty healthy males. Six marker positions along the spine's contour in each image were identified for measuring inclination, thoracic kyphosis, and lumbar lordosis angles. We estimated three inflection points around the neck, hip, and between the neck and hip, followed by identifying two adjacent marker positions per inflection point to compute its tangent. The angular deviation of each tangent line from the horizontal was computed to measure inclination angles. Thoracic kyphosis and lumbar lordosis angles were computed by the angular difference between the two adjacent tangents. A deep neural network was trained with 500,000 iterations using the labeled images from 18 participants (388 and 44 images for training and test set) and then evaluated using the unseen images (2 participants, 48 images; evaluation set).

Findings: The mean total training and test errors were <2 pixels (∼ 0.6 cm). The total error in the evaluation set was qualitatively comparable (∼ 3 pixels = ∼ 0.9 cm), suggesting the model performance was maintained in the unseen data. The angle values between labeled and network-predicted marker positions were similar in the evaluation set.

Interpretation: The network training with a relatively small number of images was successful based on the small error values observed in the evaluation set. The model may be an affordable, automated, and non-contact measurement tool for the human spine curvature.

Keywords: DeepLabCut; Kyphosis; Lordosis; Lumbar; Thoracic.