Development of automated system for computing CVI for assessment of vascular status of the choroid

Year Submitted: 2025
Published Date: 08 April 2025

Tags: Care Process ＆ Redesign, Quality Improvement, Clinical Practice Improvement, Workflow Redesign, Technology, Digitalisation, Automation, Machine Learning, Data Modelling, Data Analytics, Data Management

About this Content

Aims

The project aims to develop a fully automated customized software to analyze OCT scans and to derive important and robust markers for retino-choroidal structural changes for diagnosis, disease stratification, prognostication and prediction of treatment response. It also aims to characterize various markers for the choroidal structure in patients with systemic and ocular diseases based on OCT scans. In particular, choroidal vascularity index (CVI), which a novel OCT-derived choroidal marker, will be computed using a customized automated software and will be correlated with diagnosis (disease vs. non-disease), disease stratification (sub-phenotyping of disease), prognostication (the outcome of disease) and prediction of treatment response.

Background

The choroid is an intricate organ with high complexity and poor physical accessibility, current study methods are suboptimal. Histopathological methods utilize post-mortem samples and do not enable quantitative, repeatable assessments in living patients. Other modalities like ultrasonography (US), fundus fluorescein angiography (FFA) and indocyanine green angiography (ICGA) are not ideal to provide an objective and quantitative depiction of the choroid’s vasculature. They have suboptimal resolution for cross sectional analysis of a highly complex structure like the choroid. Moreover, FFA and ICGA are not practical tools for follow ups due to their invasive nature and possible complications from intravenous injection of dyes and possibility of an extravasation of the fluorescein dye from the highly fenestrated choriocapillaris.

Comparing against other biomarkers, OCT based parameters have high resolution and are less operator dependent and non-invasive. The technology improved over time, developing ‘time-domain’, ‘spectral-domain’, ‘swept source’ and ‘enhanced depth imaging’ modes, further increasing resolution and reproducibility. Such advancement in technology fuelled a rise in research interest for choroid biomarker. Choroidal thickness (CT) was among the first OCT based surrogate markers to be proposed, which is a measurement from the posterior edge of the hyperreflective RPE to the choroid scleral interface. However, CT varies significantly with patient physiological variables. Thus, our project proposes CVI as a robust and reproducible biomarker.

Many published studies based on CVI have small sample sizes and short follow-ups as the ability to analyse huge sample size is largely limited by the tedious manual analysis of OCT images. Thus, our project is focused on developing an automated algorithm that generate accurate estimates of CVI for both individual B scans and volume scans to overcome the problem. This project will establish CVI’s clinical relevance b

Methods

Problem 1: Automated demarcation of the choroid vasculature is not always accurate.
Solution 1: We created a Comprehensive Ocular Imaging Network (COIN) Consortium Agreement and invited global researchers to collaborate on a standardized platform to pool together large sets of robust de-identified data and images for machine learning. We also manually segmented images which demarcation are not as accurate to assist in machine learning.

Problem 2: The interface of the developing software or COIN portal was not user-friendly and impractical, and the de-identification of the images were highly dependent on the integrity of the collaborators.
Solution 2: We constantly give feedback to the vendor through a weekly online meeting for troubleshooting and adjustment. For example, multiple clicks were initially needed to expand the folder extension before reaching the uploaded image for analysis. This is fairly time-consuming given the hundreds and thousands of images to review. An overview of a branched project folder was then suggested for ease of navigation with only a few clicks. We also devised an automated de-identification system in the software to prevent collaborators from uploading files with identifiers so as to comply with personal data protection. The mapping of the de-identified codes to the identifiers is only retrievable by the investigators of the specific project.

Problem 3: The values of the total choroidal area (TCA), luminal area (LA) and stromal area (SA) reported are dependent on the image resolution exported from the machine, the image size read when imported into the automated software versus a manual image processing programme.
Solution 3: We standardized the image units in the software to pixels and enable the algorithm to compute the image resolution at different axes and dimensions at different magnification and span. These can then be used to calculate the true value or representation of the choroid regardless of the imaging machine used to

Results

There is a 50% increase in the accuracy of the auto-segmentation of the choroid vasculature and computation of CVI. There is a reduction by 60% in the time spent on navigating through the project images on the COIN portal and data extraction. This allows clinical investigator to focus on clinical duties or other research projects while the automated algorithm runs through the analysis, which is a big step towards a full automation of the software with minimal reliance on human assistance. Large batches of images can also be analyzed with half the amount of time spent on segmentation and binarization of images using a manual image processing programme. This enables more follow ups, especially for studies on diurnal variation.

Lessons Learnt

In a fast-paced clinical setting, subject recruitment is a challenge. A subject was once recruited even though he failed to meet one of the eligibility criteria. As a result, a non-compliance report was filed to the ethics board. To prevent the recurrence of similar mistake, we created an eligibility checklist for subject screening to ensure compliance to the eligibility criteria before enrollment into the study. We learnt to be more meticulous and organized when screening for patient’s eligibility to join the study after seeking their informed consent.

In order to minimize clinical disruption, especially amidst a pandemic where we have to comply with more regulations implemented, we had arranged meetings with a few representatives of the different groups of healthcare professionals, such as the optical technicians, the nurses and the patient service associates to discuss the workflow of our project. The representatives will then disseminate the information to each of their family group. In this, we learnt the importance of effective communication and teamwork to ensure that the parties involved are on the same page the continuation of research activities

Keywords

Ocular inflammation, ocular infection, novel biomarker, machine learning, artificial intelligence

Innovators' Details

Connect with this contributor!

Adj A/Prof Rupesh Agrawal - rupesh_agrawal@ttsh.com.sg

Project Attachment

CVI.pdf