Using Deep Learning for the Classification of Images Generated by Multifocal Visual Evoked Potential

Multifocal visual evoked potential (mfVEP) is used for assessing visual functions in patients with pituitary adenomas. Images generated by mfVEP facilitate evaluation of visual pathway integrity. However, lack of healthy controls, and high time consumption for analyzing data restrict the use of mfVEP in clinical settings; moreover, low signal-noise-ratio (SNR) in some images further increases the difficulty of analysis. I hypothesized that automated workflow with deep learning could facilitate analysis and correct classification of these images. A total of 9,120 images were used in this study. The automated workflow included clustering ideal and noisy images, denoising images using an autoencoder algorithm, and classifying normal and abnormal images using a convolutional neural network. The area under the receiver operating curve (AUC) of the initial algorithm (built on all the images) was 0.801 with an accuracy of 79.9%. The model built on denoised images had an AUC of 0.795 (95% CI: 0.773-0.817) and an accuracy of 78.6% (95% CI: 76.8-80.0%). The model built on ideal images had an AUC of 0.985 (95% CI: 0.976-0.994) and an accuracy of 94.6% (95% CI: 93.6-95.6%). The ensemble model achieved an AUC of 0.908 and an accuracy of 90.8% (sensitivity: 94.3%; specificity: 87.7%). The automated workflow for analyzing mfVEP plots achieved high AUC and accuracy, which suggests its possible clinical use.