A Performance Evaluation Method for Unsupervised OCT Phenotype Discovery using Deep Learning

Abstract

Purpose : We have previously demonstrated the utility of deep unsupervised phenotype discovery with OCT imaging for genetic analysis [1]. However, this method suffers from uncertainty in the number of patterns and ground truth for performance evaluation. We propose an unsupervised performance evaluation method that sets a criterion for clustering performance evaluation and can be used for determining the number of patterns.

Methods : We developed and trained two deep learning (DL) models using an autoencoder that learns the feature of OCT macular images in a pixel-wise manner and an unsupervised representation learning-based method which maps OCT macular images to feature-wise 128-element vectors with data augmentation and contrastive learning without human annotation. After data reduction with UMAP and using the probabilistic model selection method (AIC/BIC), the numbers of patterns is defined for the dataset for the output embedding of the trained DL models. Next, clustering is performed with the Gaussian Mixture Model (GMM). Finally, was use correlation analysis for images within a cluster and images in different clusters in both pixel space and feature space.

Results : We used 86115 images from UK biobanks (UKBB), 17722 images from LIFE, and 8436 images from the Mass Eye and Ear (MEE) datasets. From each dataset, we used two types of images: retinal nerve fiber layer (RNFL) and ganglion cell-inner plexiform layer (GCIPL) contour maps. Correlation analysis in the pixel space did not provide meaningful results (Fig. 1). However, correlation analysis in the feature space domain showed promising results demonstrating a stronger correlation between the images within clusters compared to images in different clusters (Fig. 2). We compared the results with the autoencoder DL model and found that the correlation in pixel-wise clustering is less than the feature-wise clustering (Fig. 2). Comparing patterns among 3 datasets, we found some patterns in one dataset were not found in another dataset (the withe rows in the Fig. 2).

Conclusions : Here we propose a criterion for validating the optimal number of OCT-based patterns obtained using unsupervised deep learning. Importantly we demonstrate our phenotypes can be represented in external datasets, reinforcing their potential utility for future gen etic discovery and risk prediction.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.