Multistream Deep Learning Models Using Multimodal Optical Coherence Tomography for Predicting Visual Impairment in...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Epiretinal membrane (ERM) can cause visual impairment.
• Accurate prediction of visual impairment is crucial for managing ERM.
• Optical coherence tomography (OCT) provides detailed retinal imaging.

Purpose of the Study:

• To develop multistream deep learning models using multimodal OCT images to predict visual impairment in ERM.
• To identify OCT imaging features that serve as biomarkers for visual impairment in ERM.

Main Methods:

• Retrospective enrollment of idiopathic ERM patients.
• Collection of eight types of OCT images: B-scan, en face OCT angiography, and retinal thickness maps.
• Development of multistream deep learning models for visual impairment prediction.
• Utilized Grad-CAM for heatmap visualization.

Main Results:

• Single-stream models showed variable performance, decreasing in external validation.
• Multistream models with two or three inputs improved predictive performance.
• An eight-stream model integrating all modalities achieved 90.90% accuracy in development and 80.00% in external validation.
• Heatmaps highlighted foveal/parafoveal areas and retinal changes as key prediction indicators.

Conclusions:

• Multimodal OCT imaging, including B-scan, en face OCT angiography, and retinal thickness maps, can predict visual impairment in ERM using deep learning.
• The multistream deep learning approach enhances predictive accuracy.
• This method may help localize critical retinal regions affected by ERM, aiding in understanding visual compromise.