Detection of Abnormal Extraocular Muscles in Small Datasets of Computed Tomography Images Using a Three–dimensional Variational Autoencoder: A Pilot Study

Abstract

We sought to establish a deep learning-based unsupervised algorithm with a three–dimensional (3D) variational autoencoder model (VAE) for the detection of abnormal extraocular muscles that are difficult to annotate in small datasets of orbital computed tomography (CT) images. 276 CT images of normal orbits were used for model training; 58 CT images of normal orbits and 96 of abnormal orbits (with extraocular muscle enlargement caused by thyroid eye disease) were used for validation. A VAE with a 3D convolutional neural network (CNN) was developed and trained for anomaly detection. All images were preprocessed to emphasize extraocular muscles and to suppress background noise (e.g., high signal intensity from bones) during model training. Model validation was conducted with normal and abnormal validation CT datasets not used for model training. The optimal cut-off value was identified through receiver operating characteristic (ROC) curve analysis. The ability of the model to detect muscles of abnormal size was assessed by visualization of differences between input and output images. During the training epochs, the 3D VAE model did not exhibit overfitting. During validation with normal and abnormal datasets, the model achieved an area under the ROC curve of 0.804, sensitivity of 87.9%, specificity of 72.9%, accuracy of 78.6%, and F1-score of 0.809. Abnormal CT images correctly identified by the model showed differences in extraocular muscle size between input and output images. The proposed 3D VAE model showed potential to detect abnormalities in small extraocular muscles using a small dataset, similar to the diagnostic approach used by physicians. Unsupervised learning can serve as an alternative detection method for medical imaging studies in which annotation is difficult or impossible to perform.