Screening of Moyamoya Disease from Retinal Photographs: Development and Validation of Deep Learning Algorithms

Abstract

Background: Moyamoya disease (MMD) is a rare and complex pathological condition characterized by an abnormal collateral circulation network in the basal brain. This disease's diagnosis and progression are unpredictable and influenced by many factors. MMD is usually diagnosed and monitored by angiographic studies. However, timely and appropriate management of patient care is limited by the frequency of these tests, the complications associated with conventional angiography, and their high cost and time requirements (especially for MRA and CTA). The blood vessels that supply the eyes can also be affected by the disease, manifesting as a range of ocular symptoms. Therefore, a deep learning model was developed and verified using real-world data to determine the occurrence and severity of the disease using retinal photographs.

Methods: In this observational cohort study, retinal photographs were gathered and analyzed with the aim of constructing a deep learning model for the screening and assessing of severity of MMD. From 2006 to 2022, at Severance Hospital in South Korea, 1850 retinal photos from 155 moyamoya patients and 3835 from 1649 healthy participants were collected. The model was evaluated based on the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and f1-score. Furthermore, image-based attention maps, progressive erasing, and progressive restoration (PEPPR) were performed to validate the robustness of the model.

Findings: A dataset of 322 retinal photographs from 67 patients with MMD and 3835 retinal photographs from 1649 healthy individuals was used to develop a screening and severity prediction model for the disease. The average age of the patients with MMD was 44⋅1 years, and the average follow-up time was 115 months. Stage 2 (36⋅6%) was the most common, followed by three, four, one, zero, and six. The moyamoya screening model had an average AUC of 94⋅6, with the best cutoff point having 90⋅36 sensitivity and 89⋅76% specificity. Among the moyamoya severity prediction models, most had an average AUC of ≥78%, with stage 3 showing the highest performance (93⋅6%). Attention map and PEPPR identified the fundus' vascular region as an important region for deep learning model classification. Our severity prediction model had an AUC of 70% when only 50% of the important regions were remained, and the top 10% of the important regions highlighted main vascular portion of the fundus.

Interpretation: This study showed that retinal photographs could be used as MMD screening and severity biomarkers. The severity of the disease was classified by a deep learning algorithm that performed comparably to that of MRA. Retinal photos may be novel biomarkers for MMD diagnosis and progression.

Funding: Ministry of Trade, Industry & Energy, South Korea.

Declaration of Interest: The authors disclose no conflicts.

Ethical Approval: All experiments were performed in accordance with the ethical principles of the Declaration of Helsinki[12].We followed the Strengthening the Reporting of Observational studies in Epidemiology statement[13]. This study was approved by the institutional review board of the Severance Hospital (IRB no. 2020-0473-0001). The need for informed consent was waived by the ethics committee, as this study utilized routinely collected data that were anonymously managed at all stages, including data cleaning and statistical analyses.

Keywords: Moyamoya disease, Retina, Deep learning, Screening, Severity detection