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

Area of Science:

• Ophthalmology and Neuroscience
• Medical Imaging Analysis
• Artificial Intelligence in Healthcare

Background:

• Multifocal visual evoked potential (mfVEP) is crucial for assessing visual functions in pituitary adenoma patients.
• Current mfVEP analysis is limited by a lack of healthy controls, time consumption, and low signal-to-noise ratio (SNR) in images.
• Automated analysis using deep learning could overcome these limitations.

Purpose of the Study:

• To develop and evaluate an automated deep learning workflow for analyzing mfVEP images.
• To improve the efficiency and accuracy of classifying normal versus abnormal mfVEP images.
• To assess the clinical utility of the automated workflow.

Main Methods:

• A dataset of 9,120 mfVEP images was utilized.
• An automated workflow involved image clustering, denoising via autoencoder, and classification using a convolutional neural network (CNN).
• An ensemble model combined results for improved performance.

Main Results:

• The initial algorithm achieved an AUC of 0.801 and accuracy of 79.9%.
• The model trained on denoised images showed an AUC of 0.795 and accuracy of 78.6%.
• The model trained on ideal images achieved a high AUC of 0.985 and accuracy of 94.6%.
• The ensemble model demonstrated excellent performance with an AUC of 0.908 and accuracy of 90.8%.

Conclusions:

• The automated deep learning workflow significantly enhances the analysis of mfVEP images.
• The high accuracy and AUC suggest the potential for clinical application in diagnosing visual pathway impairments.
• Deep learning offers a promising solution for overcoming current challenges in mfVEP analysis.