2D Photogrammetry Image of Scoliosis Lenke Type Classification Using Deep Learning

Abstract

Scoliosis is a spine deformity where 40% to 80% of victims are among adolescence. The current clinical approach uses radiographic images to determine the type of spine deformity. However, the adolescence is in the risk of cancer due to high exposure to radiation. A non-invasive approach using 2D-photogrammetric approach has been introduced and proven to be the best tool to classify scoliosis Lenke type. Convolution Neural Network (CNN) has been examined to classify either Lenke type 1 or other types of Lenke. Experiments were conducted to determine the relationship between different sizes of an input image and the best number of layers to the Lenke classification accuracy. Forty-five (45) numbers of 2D-photogrammetric images have been collected for these experiments and the results show that the three (3) convolve layers with the filter size 3x3 achieve higher accuracy compared to other filter sizes. It also shows that the image size is crucial to achieving higher accuracy and our investigation indicates that 180x180 input size leads to the high accuracy of 84.6%.