An efficient ensemble method for detecting spinal curvature type using deep transfer learning and soft voting classifier

Abstract

Scoliosis is a spinal deformity that negatively affects the body's distribution of weight in several ways. This deformity can be classified as C-shaped or S-shaped. These deformities balance each other at different times, which makes their types undistinguishable. As part of the development of spinal deformity evaluation systems, machine learning algorithms are utilized to evaluate spinal deformities to assist both patients and doctors. There are many types of machine learning algorithms, but deep neural networks can extract features automatically. As a result, they are commonly used for analyzing medical images. A dataset of 1000 private anterior-posterior spine radiographs from Shafa Hospital in Tehran, Iran, was collected and used for this study. Due to the small size of the dataset, deep transfer learning has been employed and the accuracy of pre-trained MobileNetV2, pre-trained Xception, pre-trained ResNet152, pre-trained InceptionV3, and pre-trained DenseNet121 networks have been evaluated. Accordingly, a new approach has been developed based on voting ensemble classifiers together with five pre-trained networks. As a final note, the voting ensemble model which improved spine curvature classification accuracy by about 7% for the test set can be used as a model to develop spine deformity evaluation systems that would assist radiologists, physicians, and orthopedists in evaluating spine deformities.