Accurate automated Cobb angles estimation using multi-view extrapolation net

Abstract

Accurate automated quantitative Cobb angle estimation that quantitatively evaluates scoliosis plays an important role in scoliosis diagnosis and treatment. It solves the problem of the traditional manual method, which is the current clinical standard for scoliosis assessment, but time-consuming and unreliable. However, it is very challenging to achieve highly accurate automated Cobb angle estimation because it is difficult to utilize the information of Anterior-posterior (AP) and Lateral (LAT) view X-rays efficiently. We therefore propose a Multi-View Extrapolation Net (MVE-Net) that provides accurate automated scoliosis estimation in multi-view (both AP and LAT) X-rays. The MVE-Net consists of three parts: Joint-view net learning AP and LAT angles jointly based on landmarks learned from joint representation; Independent-view net learning AP and LAT angles independently based on landmarks learned from unique independent feature of AP or LAT angles; Inter-error correction net learning a combination function adaptively to offset the first two nets' errors for accurate angle estimation. Experimental results on 526 X-rays show 7.81 and 6.26 Circular Mean Absolute Error in AP and LAT angle estimation, which shows the MVE-Net provides an accurate Cobb angle estimation in multi-view X-rays. Our method therefore provides effective framework for automated, accurate, and reliable scoliosis estimation.

Keywords: Cobb angles; Error estimation; Multi-Task learning; Scoliosis; Spinal curvature.