OCT Super-Resolution for Data Standardization using AI: A MACUSTAR report

Abstract

Purpose : In optical coherence tomography scans (OCTs) from multicenter studies, there is often large variability in image quality and resolution between scans. This impairs the consistency of biomarker quantification within studies, but also between different datasets. The aim of this study was to validate a super-resolution approach based on artificial intelligence (AI) for improving the resolution of OCTs (by increasing the density of the scan pattern) to consistently enhance data within studies to high-quality standards.

Methods : The MACUSTAR cohort, a European multicentre study, was used as a training set with 743 OCTs from 181 patients and validation set with 26 OCTs from 26 patients (n=3 no AMD, n=2 early AMD, n=18 intermediate AMD, n=3 late AMD). All scans were Heidelberg Spectralis OCTs with 241 B-scans. We trained a 3D diffusion model to generate high-resolution OCTs, which was used during evaluation to produce OCTs with 241 B-scans from OCTs with 120 B-scans. The performance was calculated using the mean squared error (MSE) on OCT volume-level between the generated B-scans and the original B-scans.

Results : The MSE between the generated B-scans from the low-resolution OCTs and the original B-scans from the high-resolution OCTs was 0.006 ± 0.004 (mean ± SD). Fig. 1 shows visual examples of the generated OCTs compared to the original B-scans in the validation set.

Conclusions : We showed the feasibility of the proposed approach to generate super-resolution OCTs, which is one of the required steps to standardize high-quality OCTs within multicenter studies. In extensions of this approach, coherence between the OCT and other modalities, such as en face imaging and other metadata, could be introduced, allowing the AI model to make better informed generative decisions.

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

 

Fig. 1: Qualitative examples of generated OCTs compared to original OCTs, along with OCT-level MSEs. In each example, the top row shows the original OCTs and the bottom row shows the generated OCTs. The first column shows the en face projection of the OCT scan, obtained by averaging all A-scans. Black and white lines next to each B-scan in the en face indicate whether that B-scan was generated or present in the low-resolution OCT, respectively. The last three columns show the B-scans of which the positions in the OCT volume are indicated in red in the en face projection.