From Machine to Machine: An OCT-trained Deep Learning Algorithm for Objective Quantification of Glaucomatous Damage in Fundus Photographs (1810.10343v1)

Abstract: Previous approaches using deep learning algorithms to classify glaucomatous damage on fundus photographs have been limited by the requirement for human labeling of a reference training set. We propose a new approach using spectral-domain optical coherence tomography (SDOCT) data to train a deep learning algorithm to quantify glaucomatous structural damage on optic disc photographs. The dataset included 32,820 pairs of optic disc photos and SDOCT retinal nerve fiber layer (RNFL) scans from 2,312 eyes of 1,198 subjects. A deep learning convolutional neural network was trained to assess optic disc photographs and predict SDOCT average RNFL thickness. The performance of the algorithm was evaluated in an independent test sample. The mean prediction of average RNFL thickness from all 6,292 optic disc photos in the test set was 83.3$\pm$14.5 $\mu$m, whereas the mean average RNFL thickness from all corresponding SDOCT scans was 82.5$\pm$16.8 $\mu$m (P = 0.164). There was a very strong correlation between predicted and observed RNFL thickness values (r = 0.832; P<0.001), with mean absolute error of the predictions of 7.39 $\mu$m. The areas under the receiver operating characteristic curves for discriminating glaucoma from healthy eyes with the deep learning predictions and actual SDOCT measurements were 0.944 (95$\%$ CI: 0.912- 0.966) and 0.940 (95$\%$ CI: 0.902 - 0.966), respectively (P = 0.724). In conclusion, we introduced a novel deep learning approach to assess optic disc photographs and provide quantitative information about the amount of neural damage. This approach could potentially be used to diagnose and stage glaucomatous damage from optic disc photographs.

• The paper demonstrates that an OCT-trained deep learning algorithm objectively quantifies glaucomatous damage in fundus photographs by using SDOCT data as a reference.
• The methodology employs a ResNet34 architecture trained on 32,820 photo-SDOCT pairs, achieving a correlation coefficient of r = 0.832 and a mean absolute error of 7.39 μm.
• Implications include enhanced glaucoma screening and diagnosis through a scalable, cost-effective alternative to subjective human assessments.

Objective Quantification of Glaucomatous Damage Using an OCT-trained Deep Learning Algorithm

The paper "From Machine to Machine: An OCT-trained Deep Learning Algorithm for Objective Quantification of Glaucomatous Damage in Fundus Photographs" explores a promising methodology in the automated assessment of glaucoma through the use of a deep learning convolutional neural network trained on spectral-domain optical coherence tomography (SDOCT) data. It presents an alternative pathway for evaluating optic disc photographs, targeting objective quantification of structural damage without relying on human labeling, which historically has been subjective and marred by inconsistency.

Overview of Research

Traditionally, the training of algorithms to detect glaucomatous damage in fundus photographs has relied heavily on subjective assessments made by expert graders. This paper seeks to circumvent the limitations inherent in human evaluations by utilizing objective SDOCT data as a reference standard for training purposes. The paper leverages a dataset comprised of 32,820 pairs of optic disc photos and corresponding SDOCT RNFL scans collected from a diverse sample of 2,312 eyes. Utilizing ResNet34 architecture, the deep learning algorithm was adeptly trained to predict average RNFL thickness from optic disc photographs, offering continuous, quantitative predictions of structural neural damage.

Key Findings

The results documented in the paper are notable for the neural network's competency in mirroring SDOCT predictive capabilities. A comparative analysis between the algorithmic predictions and actual SDOCT measurements revealed a strong correlation coefficient (r = 0.832), indicating reliability in its predictions. The mean absolute error of the algorithm's predictions stood at 7.39 μm, showcasing remarkable accuracy. Importantly, the area under the ROC curve (AUC) for discriminating between glaucoma and healthy eyes based on these predictions was 0.944 compared to 0.940 for true SDOCT measurements, signifying substantial alignment in diagnostic precision.

The paper also details the deep learning model's performance concerning the classification of images into categories of normal or abnormal, laying emphasis on its accuracy, which reached 83.7% when compared to SDOCT normative classifications.

Implications and Future Directions

The introduction of this deep learning framework represents a significant stride forward in automating the assessment of glaucomatous damage in fundus photographs. Optic disc photographs offer a cost-effective and minimally invasive means to document eye health. The presented approach could substantially enhance glaucoma screening efforts, especially in resource-constrained settings with limited access to advanced SDOCT equipment. This model may serve as a cornerstone for developing scalable, portable solutions for early diagnosis and ongoing monitoring of glaucoma progression.

Looking ahead, opportunities exist to refine the algorithm further by integrating more nuanced SDOCT data, possibly utilizing sectoral measurements to capture finer detail. Exploration of these avenues could bolster the algorithm's applicability across diverse clinical scenarios. Additional longitudinal studies are necessary to validate its efficacy in detecting and monitoring disease progression, which remains crucial for impactful medical decision-making in glaucoma management.

In summary, this research marks a pivotal contribution towards the objective quantification of glaucoma from optic disc photographs and serves as a foundation for expanding the role of deep learning in ophthalmology's diagnostic arsenal.