Handcrafted and Deep features based Classification of Scoliosis

Abstract

The Spinal cord acts as the central transmission line connecting the Brain with all other body organs. Vertebrae are 33 uneven bones stacked over one another that holds the whole skeleton structure. Scoliosis is the three-dimensional spinal deformity which commonly occurs during the growing age and erupts before puberty. It is further classified in two Shapes C and S. Our research work has two stages, in first stage we segment out the vertebral column using Mask-RCNN. The segmented column is used for features extraction and in stage two feature based classification is done for normal, C and S shape of scoliosis using AASCE2019 dataset. A comparative study on multiple image classification networks is also conducted and based on results EfficientNet-B4 is selected for formulation of hybrid feature set. The accuracy achieved using Random forest classifier, for handcrafted and deep features was up to 94.32% and 89.66%. Hybrid feature set formulated with combination of deep and handcrafted features attained accuracy up to 94.45%.