Automated Segmentation and Characterization of Retinal Hyperreflective Foci in Age-Related Macular Degeneration

Retinal hyperreflective foci (HRF) are associated with higher risk of progression to advanced age-related macular degeneration (AMD). To enable automated detection, we developed Foci-Net, a convolutional neural network for segmenting HRF in optical coherence tomography (OCT) volumes.

This cross-sectional study included eyes clinically diagnosed with intermediate to advanced AMD and healthy controls. We acquired 6 × 6-mm macular-centered scans using 2 OCT systems. Foci-Net modifies the U-Net by replacing the bottleneck with a fine-to-coarse feature extraction block to improve segmentation of both small and large foci. HRF volume was quantified as the proportion of total volume within each retinal layer and characterized by anatomic location. Model performance was evaluated at the count, pixel, eye, and B-scan levels using F1-score, area under the curve (AUC) of the receiver operating characteristic (ROC) curve, precision, sensitivity, and specificity.

Sixty-one volumetric OCT scans (50 AMD and 11 healthy control eyes) obtained from 50 participants, age 77.65 ± 9.24 years (mean ± SD), of whom 72% were women. Foci-Net achieved F1-scores of 73.0 ± 3.8% counts (based on en face analysis) and 71.6 ± 29.1% pixels (based on cross-sectional analysis) level relative to the ground truth in 3-fold cross-validation. Diagnostic performance was strong, with an AUC of 100 ± 0.0% (eye) and 92.7 ± 0.3% (B-scan) level. HRF was least (2.10%) abundant in the ganglion cell complex and most (55.37%) in the outer nuclear layer.

A deep learning approach enables fast HRF segmentation, supporting its role as a biomarker for predicting AMD progression.

The proposed HRF segmentation method could enhance AMD diagnosis and prognosis.