When Two Eyes Don't Suffice-Learning Difficult Hyperfluorescence Segmentations in Retinal Fundus Autofluorescence...

Area of Science:

• Ophthalmology
• Medical Imaging
• Computational Biology

Background:

• Hyperfluorescence (HF) and reduced autofluorescence (RA) are key biomarkers in fundus autofluorescence (FAF) imaging for assessing retinal pigment epithelium (RPE) health.
• FAF biomarkers are crucial for monitoring diseases like geographic atrophy (GA) and central serous chorioretinopathy (CSCR).
• Human annotation of FAF images shows significant inter- and intra-grader variability, particularly for biomarker boundaries, impacting reliability.

Purpose of the Study:

• To develop an automated method for segmenting HF and RA in FAF images that overcomes human annotation variability.
• To evaluate the performance of a segmentation ensemble approach against expert human graders.
• To create reliable ternary segmentations for confident biomarker identification and detection.

Main Methods:

• A segmentation ensemble model was trained using FAF images with single annotations.
• The ensemble's performance was evaluated by comparing its segmentations to those of multiple expert graders using Dice scores.
• Ternary segmentations were generated using the ensemble's mean predictions and variance to classify image areas.

Main Results:

• Human expert agreement for HF segmentation ranged from 63-80% Dice score, and for RA from 14-52%.
• The segmentation ensemble achieved expert-like agreement, with Dice scores of 64-81% for HF and 21-41% for RA.
• Ternary segmentations demonstrated high precision (97%) for confident background/HF identification and high recall (99%) for detecting all HF instances.

Conclusions:

• Automated segmentation using a ensemble approach can achieve expert-level performance in identifying HF and RA in FAF images.
• The proposed ternary segmentation method provides reliable and interpretable results, distinguishing confident biomarker regions from uncertain ones.
• This approach offers a promising tool for objective and consistent assessment of RPE health in retinal disease monitoring.