Free papers detail | 11th Asia Pacific Tele Ophthalmology Society Symposium

Title
Automated Artificial Intelligence Framework for High-Precision Quantification of Retinal Thinning and Atrophy

Authors
Kah Meng Ang, Anais Monet, Sabrina Noritake, Denise Ng, Yann Malato, Weiwei Luo

Presenting
Kah Meng Ang

PURPOSE:
Evaluating an artificial intelligence (AI) framework to quantify retinal thinning and focal atrophy in a mouse model of Geographic Atrophy (GA) using high-density three-dimensional optical coherence tomography (OCT).

METHODS:
Retinal atrophy was induced in C57BL/6JRj lasered mice model. To enhance image quality and the signal-to-noise ratio (SNR), a B-scan alignment and frame-averaging protocol was implemented, to visualize subtle features such as the ellipsoid zone (EZ). A supervised pipeline processed 1000 B-scans per volume. Addressing segmentation artifacts, we trained a model for en face-level analysis. Multi-Gaussian fits defined objective "thinning thresholds" to define Total Attenuation Zones (TAZ) for focal collapse and Partial Attenuation Zones (PAZ) for early thinning. Progression from Day 14 to Day 28 was assessed using paired t-tests.

METHODS:

Retinal atrophy was induced in C57BL/6JRj lasered mice model. To enhance image quality and the signal-to-noise ratio (SNR), a B-scan alignment and frame-averaging protocol was implemented, to visualize subtle features such as the ellipsoid zone (EZ). A supervised pipeline processed 1000 B-scans per volume. Addressing segmentation artifacts, we trained a model for en face-level analysis. Multi-Gaussian fits defined objective "thinning thresholds" to define Total Attenuation Zones (TAZ) for focal collapse and Partial Attenuation Zones (PAZ) for early thinning. Progression from Day 14 to Day 28 was assessed using paired t-tests.

RESULTS:
The framework was accurate across three independent cohorts. The image‑enhancement protocol increased confidence in outer‑retina detection and segmentation, such as the EZ band. The hybrid B‑scan/en face approach reduced artifact‑related errors. Longitudinal analysis showed significant atrophy expansion from Day 14 to Day 28 (Day 14 vs Day 28; study 2: p<0.05; study 3: p < 0.001). PAZ served as a sensitive early biomarker preceding focal collapse, with progression slopes quantifying lesion‑growth rates.

RESULTS:

The framework was accurate across three independent cohorts. The image‑enhancement protocol increased confidence in outer‑retina detection and segmentation, such as the EZ band. The hybrid B‑scan/en face approach reduced artifact‑related errors. Longitudinal analysis showed significant atrophy expansion from Day 14 to Day 28 (Day 14 vs Day 28; study 2: p<0.05; study 3: p < 0.001). PAZ served as a sensitive early biomarker preceding focal collapse, with progression slopes quantifying lesion‑growth rates.

CONCLUSIONS:
Our AI framework enables precise monitoring of GA biomarkers, leveraging image‑enhancement and hybrid segmentations to identify retinal thinning as a key precursor to atrophy and provide a reproducible method for preclinical research and clinical tele‑ophthalmology.