Objective: This study aims to investigate the capability of artificial intelligence, more precisely deep convolutional neural networks (CNN), to automatically classify between healthy controls (HC) and Parkinson’s Disease (PD) patients based on Optical Coherence Tomography (OCT) retinal images.

Background: The advent of deep CNNs has dramatically increased the classification capability of automatic models used on biomedical images, even outperforming traditional methods based on feature engineering and machine learning. Previous studies have linked alterations in the retina to neurological diseases, such as PD [1], although none have provided acceptable classification results between HC and PD using only retinal thickness parameters. No investigation so far, to the best of the author’s knowledge, has exploited the potential of CNNs to automatically classify HC and PD using solely OCT retinal layers thickness maps.

Method: A database with 94 idiopathic PD patients and 94 sex and age-matched HC was considered in this investigation. Both right and left eyes of each subject were included in the database, where left eyes were horizontally flipped to match the nasal and temporal sectors of right eyes. OCT volumetric images of the eyes were automatically segmented using the built-in software of the scanner (Heidelberg Eye Explorer 1.9.10.0) and were pre-processed (flattening and fovea centering) following the procedure described in [2]. Images with visible segmentation errors were discarded, leaving a dataset of 186 PD and 186 HC eyes. The generated ganglion cell-inner plexiform layer (GCIPL) thickness maps were then fed to a pre-trained ResNet50 [3] CNN which was modified and fine-tuned to perform the current classification task. We used a bootstrap approach of 50 replicas of randomly distributed 5 fold cross-validation to obtain statistically meaningful results.

Results: Our automatic CNN-based classification model achieved a mean AUC with a 95% confidence interval of 0.761 [0.757,0.765] using our local dataset. Furthermore, the model achieved a mean sensitivity, specificity and accuracy of 0.714 [0.706,0.722], 0.664 [0.656,0.672] and 0.690 [0.684,0.693], respectively.

Conclusion: Deep learning techniques can extract abstract and complex parameters that allow accurate classification of HC and PD using only retinal images.

References: [1] A. Murueta-Goyena et al., «Parafoveal thinning of inner retina is associated with visual dysfunction in Lewy body diseases», Mov. Disord. Off. J. Mov. Disord. Soc., vol. 34, n.o 9, pp. 1315-1324, sep. 2019, doi: 10.1002/mds.27728.
[2] D. Romero-Bascones et al., «Foveal Pit Morphology Characterization: A Quantitative Analysis of the Key Methodological Steps», Entropy, vol. 23, n.o 6, Art. n.o 6, jun. 2021, doi: 10.3390/e23060699.
[3] K. He, X. Zhang, S. Ren, y J. Sun, «Deep Residual Learning for Image Recognition», 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), jun. 2016, pp. 770-778. doi: 10.1109/CVPR.2016.90.

« Back to 2022 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/deep-learning-applied-to-retinal-oct-images-to-differentiate-parkinsons-disease-patients-from-healthy-controls/