Abstract
Purpose :
The purpose of this study was to evaluate low-dimensional feature representations of a deep convolutional neural network (CNN) designed to diagnose plus disease in retinopathy of prematurity (ROP). We used two different feature dimensionality techniques, t-stochastic neighbor embedding (t-SNE) and uniform manifold approximation and projection (umap) clustering to evaluate whether images of similar disease severity cluster together in high dimensional space, and whether change in high dimensional space may correlate with clinically observed changes in disease severity.
Methods :
We previously reported the development and evaluation of a deep CNN designed to classify plus disease in ROP. All infants received digital fundus imaging (RetCam, Natus Medical Incorporated, Pleasanton, CA) as well as ophthalmoscopic examination, and were assigned a reference standard diagnosis (RSD) using methods previously published. We analyzed 5511 posterior pole images from the Imaging and Informatics in ROP (i-ROP) dataset. In the t-SNE analysis, we compared t-SNE coordinates with RSD for plus and stage. In addition, a subset of the data (131 images from 46 patients) was extracted, gathered and divided into four cohorts based on the RSD observed over multiple visits: no progression, progression to plus, progression to APROP, and regression after bevacizumab treatment. We analyzed umap coordinates of the images at baseline and over time in the four cohorts, relative to the entire dataset.
Results :
Images from the entire dataset were mapped into t-SNE space (Figure 1) and umap space (Figure 2), and demonstrate a continuum of disease severity from no plus to pre-plus to plus in both clustering algorithms. Figure 1 demonstrates the co-clustering of eyes with stage 3 and pre-plus and plus disease. Figure 2 demonstrates changes in umap coordinates corresponding with changes in the RSD.
Conclusions :
t-SNE and umap clustering appear to reveal clinically useful information, both at a given point in time, and in changes over time. This work is broadly relevant to image-based deep learning classification systems and prognostic risk modeling using images. Future work using these dimensionality reduction techniques may reveal additional clinically relevant information and disease phenotypes.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.