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

Title
SELENA+ (Singapore Eye Lesion Analyser +) AI as Pre-Processing Tool for Research Grading: A Pilot Longitudinal Workflow Efficiency Study

Authors
Wong Chong En, Haslina Binte Hamzah, Kayathri D/O Jaya Paul, Ho Jin Yi, Gavin Tan Siew Wai

Presenting
Wong Chong En

PURPOSE:
To conduct a pilot evaluation of SELENA+ (Singapore Eye Lesion Analyser +) artificial intelligence as a pre-processing tool to improve workflow efficiency in research grading, reducing trained reader time while maintaining diagnostic accuracy in longitudinal diabetic retinopathy studies.

METHODS:
Pilot prospective research study of diabetic patients with fundus photographs obtained at baseline, 6 months, and 12 months. Images underwent both SELENA+ AI pre-processing (30 seconds per eye) and manual grading by a trained reader (4 minutes per eye) using ETDRS severity scale. Referability was defined as ≥ETDRS level 35.

RESULTS:
Of 41 baseline patients (82 eyes), 39 patients (78 eyes, 95.1%) completed 6-month follow-up and 27 patients (54 eyes, 65.9%) completed 12-month follow-up. SELENA+ demonstrated perfect sensitivity (100%) for identifying cases requiring trained reader review at all timepoints, maintaining zero missed referrals while reducing routine grading workload. Specificity was 75.9%, 100%, and 93.8% at baseline, 6 months, and 12 months respectively, allowing the trained reader to focus on the 24% of cases requiring detailed assessment. Agreement was 82.9%, 100%, and 96.3% respectively. All eyes progressing to referrable disease were correctly flagged for reader review. The AI pre-processing achieved 87.5%-time reduction for non-referrable cases, reducing per-eye assessment from 4 minutes to 30 seconds.

RESULTS:

Of 41 baseline patients (82 eyes), 39 patients (78 eyes, 95.1%) completed 6-month follow-up and 27 patients (54 eyes, 65.9%) completed 12-month follow-up. SELENA+ demonstrated perfect sensitivity (100%) for identifying cases requiring trained reader review at all timepoints, maintaining zero missed referrals while reducing routine grading workload. Specificity was 75.9%, 100%, and 93.8% at baseline, 6 months, and 12 months respectively, allowing the trained reader to focus on the 24% of cases requiring detailed assessment. Agreement was 82.9%, 100%, and 96.3% respectively. All eyes progressing to referrable disease were correctly flagged for reader review. The AI pre-processing achieved 87.5%-time reduction for non-referrable cases, reducing per-eye assessment from 4 minutes to 30 seconds.

CONCLUSIONS:
This pilot study demonstrates that SELENA+ AI effectively serves as a pre-processing triage tool for research grading, safely identifying routine cases while ensuring perfect sensitivity for cases requiring trained reader assessment. The consistent workflow efficiency gains across longitudinal timepoints provide proof-of-concept for larger-scale validation studies in diabetic retinopathy research.

CONCLUSIONS:

This pilot study demonstrates that SELENA+ AI effectively serves as a pre-processing triage tool for research grading, safely identifying routine cases while ensuring perfect sensitivity for cases requiring trained reader assessment. The consistent workflow efficiency gains across longitudinal timepoints provide proof-of-concept for larger-scale validation studies in diabetic retinopathy research.