JMIR Publications

ABSTRACT

Background:

Brolucizumab is an effective treatment for neovascular age-related macular degeneration (nAMD). However, its use has been closely associated with the occurrence of intraocular inflammation (IOI), which may lead to severe and potentially irreversible visual impairment. These safety concerns have limited the clinical use of brolucizumab. Therefore, a reliable method for detecting IOI following brolucizumab injection is urgently needed.

Objective:

To evaluate the feasibility of quantitatively detecting IOI-related features following intravitreal brolucizumab injection in patients with nAMD using optical coherence tomography (OCT)-based deep learning analysis.

Methods:

This study retrospectively enrolled treatment-naïve nAMD patients who developed IOI after brolucizumab injection from Kyoto University Hospital and Gunma University Hospital between 2020 and 2025. Kyoto University Hospital provided 11 eyes of 11 patients with IOI, and Gunma University Hospital provided 20 eyes of 20 patients with IOI and 10 eyes of 10 patients without IOI. We developed a novel framework for longitudinal assessment of IOI by explicitly detecting and quantifying subtle IOI-related features at the object level. The detection model was trained on 304 OCT images from 11 patients at Kyoto University Hospital using patient-level ten-fold cross-validation to prevent data leakage. Externally validation was subsequently performed on 30 patients at Gunma University Hospital. To improve micro-target detection, high-resolution layers and a lightweight channel purification module (CPM) were integrated. Two quantitative imaging indicators, the vitreous cell score (VCS) and vitreoretinal interface irregularity score (VRIIS), were automatically derived from the detected features. Their changes from baseline to post-injection were subsequently integrated into a longitudinal assessment, with an increase in either feature considered indicative of increased IOI risk.

Results:

The proposed model reduced the number of parameters from 2.59 million to 0.63 million while improving detection performance, with a 4.61 percentage points increase in precision and a 3.68 percentage points increase in recall. In the external validation dataset, the proposed model correctly identified 28 out of 30 cases, achieving an accuracy of 93.3%, compared with 73.3% (22 of 30 cases) for the reference model.

Conclusions:

This study presents a framework for the objective detection and quantification of IOI-related features in OCT images and the longitudinal assessment during nAMD treatment. The proposed model achieved improved detection performance with approximately one-quarter of the original number of parameters and demonstrated 93.3% accuracy on external validation using data from a different institution and OCT device, supporting its potential clinical applicability. Misclassified cases involved either preretinal hemorrhage or anterior segment inflammation, indicating the limitation of OCT-based assessment alone. Future work will incorporate multimodal imaging to further improve IOI detection performance in nAMD treatment.