A lightweight convolutional neural network for intraocular inflammatory cell differentiation with ultrahigh-resolution OCT

Abstract

Purpose : The differential diagnosis of uveitis is broad and challenging. The presence of cells in the aqueous or vitreous humor is a hallmark of ocular inflammation. In this study, we developed a convolutional neural network (CNN) model to classify inflammatory cell types with ultrahigh-resolution (UHR) optical coherence tomography (OCT).

Methods : Neutrophils, lymphocytes, and monocytes separated from the periphery blood of a healthy donor were suspended in phosphate-buffered saline (PBS) solution and placed in cuvettes. The cell suspensions were scanned with a retinal UHR-OCT prototype with 1.8 µm axial resolution (full-width-half-maximum in tissue). A volumetric scan pattern (1 mm × 1 mm, 1000 raster lines of 1000 axial scans each) was used. Cells were located on the OCT volume as hyperreflective spots. For each cell, a 3-dimensional sampling volume of 16×8×4 voxels (16 μm × 8 μm × 4 μm, axial × transverse × lateral), whose center matched the center of individual cells, was extracted from the OCT image. The cell images were normalized to remove the dependency on overall signal intensity information and served as input for training a bespoke lightweight CNN with two convolutional and four fully connected layers. The proposed model underwent validation via a five repetition of five-fold cross-validation (train 70%, validation 10%, and test 20%).

Results : UHR-OCT images of 1841 neutrophils, 1662 lymphocytes, and 1793 monocytes were identified for the CNN model. The uniform manifold approximation and projection (UMAP) of features from the last fully connected layer of the CNN demonstrated good clustering of cell types (Figure 1). The CNN achieved an overall accuracy of 91.6 ± 0.9% in distinguishing monocytes, neutrophils, and lymphocytes at the single-cell level, with class-specific accuracies of 90.0 ± 2.0%, 90.6 ± 2.5%, and 94.4 ± 2.1%, respectively.

Conclusions : Artificial intelligence-assisted OCT is a noninvasive way of determining the composition of cells associated with inflammation within a transparent medium. This method may be applicable to the characterization of cells within the aqueous and vitreous humor which is relevant to the diagnosis and treatment of uveitis.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.

 

Figure 1. Uniform manifold approximation and projection (UMAP) of features representing neutrophils, lymphocytes, and monocytes.