Artificial intelligence and machine learning in spine research

Abstract

Artificial intelligence (AI) and machine learning (ML) techniques are revolutionizing several industrial and research fields like computer vision, autonomous driving, natural language processing, and speech recognition. These novel tools are already having a major impact in radiology, diagnostics, and many other fields in which the availability of automated solution may benefit the accuracy and repeatability of the execution of critical tasks. In this narrative review, we first present a brief description of the various techniques that are being developed nowadays, with special focus on those used in spine research. Then, we describe the applications of AI and ML to problems related to the spine which have been published so far, including the localization of vertebrae and discs in radiological images, image segmentation, computer-aided diagnosis, prediction of clinical outcomes and complications, decision support systems, content-based image retrieval, biomechanics, and motion analysis. Finally, we briefly discuss major ethical issues related to the use of AI in healthcare, namely, accountability, risk of biased decisions as well as data privacy and security, which are nowadays being debated in the scientific community and by regulatory agencies.

Keywords: artificial neural networks; deep learning; ethical implications; outcome prediction; segmentation.

Figure 1

Schematic overview of the main branches of artificial intelligence (AI), including machine learning (ML) methods which are having an impact on spine research

Figure 2

Schematic representation of an artificial neural network (A), a deep network (B), and a unit, also called artificial neuron (C). In each unit, the inputs (“x _1,3 ”) are multiplied by weights (“w _1,3”), summed to a bias term (“+t”), and the total sum is processed by a linear or nonlinear activation function (“φ”)

Figure 3

Examples of a plausible good fitting (left), underfitting (center), and overfitting (right) in a binary classification task

Figure 4

Schematic representation of a simple support vector machine (SVM) used for binary classification. In brief, the SVM builds the optimal hyperplane (in green) which separates the two classes maximizing the gap between them. A non‐optimal hyperplane (in orange) which correctly separates the two classes, but with a smaller gap, is also shown. The SVM operates in the feature space (“x1” and “x2” in the exemplary figure)

Figure 5

Example of a decision tree trained to predict the risk of failure of pedicle screws. Reproduced with permission from Varghese et al36

Figure 6

Schematic representation of a convolutional neural network (CNN), here exemplary aimed at performing the grading of disc degeneration on T2‐weighted MRI scans based on the scheme presented by Pfirrmann et al.15 In a convolutional layer, a small filter convolves over the data creating a series of activation maps; these maps can be downsampled by pooling layers, and then processed by another convolutional layer. In the simplest forms of a CNN, one or more fully connected layers perform the final classification or regression decision

Figure 7

Examples of localization of the vertebral centroids from a literature study,61 dealing with different types of CT images (from left to right: standard, low resolution, noisy, cropped). Manual annotations by an expert operator are shown in yellow, whereas the computer predictions are in red. The numbers indicate the mean absolute error (MAE) with respect to the manual annotations. Reproduced with permission from Glocker et al61

Figure 8

Five automated segmentation methods for CT scans developed in the frame of the grand challenge organized by the International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI) Workshop on Computational Spine Imaging (CSI 2014). Reprinted with permission from Yao et al87

Figure 9

Top: workflow to perform classification tasks on lumbar MRI scans from a literature study.99 First, vertebrae are detected, then the volumes corresponding to the intervertebral discs are extracted and passed to a classifier. Bottom: the various radiological parameters (Pfirrmann grading of disc degeneration15; disc narrowing; spondylolisthesis; central canal stenosis; endplate defects; marrow changes) automatically extracted from the images in the same study. Reproduced from Jamaludin et al99

Figure 10

Eleven clusters of spine curves of patients suffering from adolescent idiopathic scoliosis, automatically determined from a large database of biplanar radiographs.110 For each cluster, exemplary radiographs, da Vinci views,111 coronal and top views of the three‐dimensional reconstructions are shown. Reproduced with permission from Thong et al110

Figure 11

Example of heatmap showing the importance of the various factors (first column) in determining an outcome, namely the risk of complications following posterior lumbar spine fusion, as predicted with machine learning (ML) techniques in a literature study.120 Reproduced with permission from Kim et al120