Abstract
Purpose :
Machine learning algorithms (ML) for the automatic detection of pathological OCT volumes based on structural, angiography OCTA and systemic metrics were trained and the performance was assessed.
Methods :
In this work 31 metrics associated with vascular function, full retinal thickness (RT), ganglion cell layer (GCL) thickness and systemic metrics were used to detect pathological eyes. Data from a total of 344 eyes/persons (208 healthy (72±8 years) and 136 (75±9 years) with pathologies) were used in this study. The imaging data were acquired using the ZEISS AngioPlex OCT Angiography (ZEISS, Dublin. CA) and the metrics were computed by the native software of the equipment. The lesions found by an ophthalmologist in the pathological eyes include drusen (38%), epiretinal membrane (52%), intraretinal fluid (19%) and hyper-refletive foci (8%).
The performance of two different logistic regression models was assessed: the first uses all the metrics and the other uses only the 5 most relevant features given by the SHAP (Shapley Additive Explanations) method. Three-fold cross-validation was used to test the performance of the models. The quantitative comparison was based on the area under the receiver operator characteristics curve (AUROC).
Results :
The five most relevant features given by the SHAP method were the minimum GCL thickness, the RT in the central subfield, the GCL thickness measured in the temporal inferior sector, the perfusion measured in the outer sector and the RT measured in the outer inferior subfield. An AUCROC of 0.76±0.03 was measured when all the features were used and an AUCROC equal to 0.75±0.03 was measured when only 5 metrics were used in the classification.
Conclusions :
Our results indicate that the metrics that involve the GCL thickness, the RT and the perfusion play a key role in the automatic detection of pathological OCTs volumes.
These kinds of methods are especially needed in the scope of screening programs where a very large amount of data is acquired in a very short time.
This is a 2021 Imaging in the Eye Conference abstract.