TY - UNPB

T1 - Retinal OCT disease classification with variational autoencoder regularization

AU - Laves, Max-Heinrich

AU - Ihler, Sontje

AU - Kahrs, Lüder A.

AU - Ortmaier, Tobias

PY - 2019

Y1 - 2019

N2 - According to the World Health Organization, 285 million people worldwide livewith visual impairment. The most commonly used imaging technique for diagnosisin ophthalmology is optical coherence tomography (OCT). However, analysis ofretinal OCT requires trained ophthalmologists and time, making a comprehensiveearly diagnosis unlikely. A recent study established a diagnostic tool based onconvolutional neural networks (CNN), which was trained on a large database ofretinal OCT images. The performance of the tool in classifying retinalconditions was on par to that of trained medical experts. However, the trainingof these networks is based on an enormous amount of labeled data, which isexpensive and difficult to obtain. Therefore, this paper describes a methodbased on variational autoencoder regularization that improves classificationperformance when using a limited amount of labeled data. This work uses atwo-path CNN model combining a classification network with an autoencoder (AE)for regularization. The key idea behind this is to prevent overfitting whenusing a limited training dataset size with small number of patients. Resultsshow superior classification performance compared to a pre-trained and fullyfine-tuned baseline ResNet-34. Clustering of the latent space in relation tothe disease class is distinct. Neural networks for disease classification onOCTs can benefit from regularization using variational autoencoders whentrained with limited amount of patient data. Especially in the medical imagingdomain, data annotated by experts is expensive to obtain.

AB - According to the World Health Organization, 285 million people worldwide livewith visual impairment. The most commonly used imaging technique for diagnosisin ophthalmology is optical coherence tomography (OCT). However, analysis ofretinal OCT requires trained ophthalmologists and time, making a comprehensiveearly diagnosis unlikely. A recent study established a diagnostic tool based onconvolutional neural networks (CNN), which was trained on a large database ofretinal OCT images. The performance of the tool in classifying retinalconditions was on par to that of trained medical experts. However, the trainingof these networks is based on an enormous amount of labeled data, which isexpensive and difficult to obtain. Therefore, this paper describes a methodbased on variational autoencoder regularization that improves classificationperformance when using a limited amount of labeled data. This work uses atwo-path CNN model combining a classification network with an autoencoder (AE)for regularization. The key idea behind this is to prevent overfitting whenusing a limited training dataset size with small number of patients. Resultsshow superior classification performance compared to a pre-trained and fullyfine-tuned baseline ResNet-34. Clustering of the latent space in relation tothe disease class is distinct. Neural networks for disease classification onOCTs can benefit from regularization using variational autoencoders whentrained with limited amount of patient data. Especially in the medical imagingdomain, data annotated by experts is expensive to obtain.

U2 - 10.48550/arXiv.1904.00790

DO - 10.48550/arXiv.1904.00790

M3 - Preprint

BT - Retinal OCT disease classification with variational autoencoder regularization

ER -