Predicting OCT biological marker localization from weak annotations.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Deep learning models have demonstrated efficacy in predicting biological markers within Optical Coherence Tomography (OCT) volumes for Age-Related Macular Degeneration (AMD) and Diabetic Retinopathy (DR).
• Accurate localization of fluid markers is crucial for diagnosing and managing retinal diseases like AMD and DR.

Purpose of the Study:

• To develop and validate a deep learning method for automatically locating Intraretinal Fluid (IRF) and Subretinal Fluid (SRF) within the Early Treatment Diabetic Retinopathy Study (ETDRS) rings using OCT B-scans.
• To improve the precision of fluid detection and localization in OCT scans for patients with AMD and DR.

Main Methods:

• A neural network was trained on 22,723 OCT B-scans from 460 eyes (433 patients) with AMD and DR, utilizing slice-level annotations for IRF and SRF.
• Outputs were mapped to ETDRS rings, and a custom loss function incorporating domain knowledge constrained the predictions.
• The model's performance was evaluated on 322 eyes (189 patients) with Diabetic Macular Edema.

Main Results:

• The method accurately predicted the presence of IRF and SRF in ETDRS rings, surpassing baseline performance, particularly in challenging cases.
• A high correlation coefficient of 0.946 was achieved for Intraretinal Fluid area prediction.
• The model demonstrated successful application to en-face marker segmentation and showed intra-scan consistency without using volumetric data during training.

Conclusions:

• The proposed deep learning approach effectively and accurately localizes IRF and SRF to ETDRS rings in OCT scans.
• This method offers a promising tool for enhanced diagnosis and management of retinal diseases like Diabetic Macular Edema.
• The model's ability to generalize to en-face segmentation suggests broader applicability in retinal image analysis.