Post

New research · Ophthalmology
Research square · 1d
AI / informaticsResearch square · 2026

Deep Learning Optimisation Strategies for Uveal Melanoma Detection Using Ultra-Widefield Photography.

Michael Heiferman, Sanjay Ganesh, Virginia Tasso … Darvin Yi
Read paper
OphthalmologyAI / informatics

Adding healthy eye images as a third class boosted uveal melanoma detection accuracy.

Deep Learning Optimisation Strategies for Uveal Melanoma Detection Using Ultra-Widefield Photography.

Michael Heiferman … Darvin Yi
Research square · 2026
Background

Uveal melanoma (uveal melanoma) is the most common primary intraocular malignancy in adults and carries significant metastatic risk.

Purpose

This study evaluates data-centric optimisation strategies for deep learning classification of uveal melanoma using ultra-widefield (ultra-widefield) fundus photography.

Methods

This retrospective study analysed ultra-widefield fundus photographs from 784 patients (864 images) seen at the University of Illinois Chicago eye clinic.

n = 864 images
Results

AI told nevus, melanoma, and healthy eyes apart with near-perfect accuracy

Baseline model
0.906AUC
Top-performing model
0.987AUC
More results

The baseline model achieved an AUC of 0.906 ± 0.032.

Dataset augmentation approaches yielded minimal performance gain, and Multi-strategy models showed no additive benefit.

“
Conclusion · 1 of 3

Data-centric optimisation significantly influences deep learning performance for uveal melanoma detection.

Conclusion · 2 of 3

Three principles emerge: healthy class addition improves specificity and anatomical feature learning; data quality may outweigh quantity; and contextual input tuning is a key model parameter.

Conclusion · 3 of 3

These findings offer a practical framework for developing clinically robust, physician-supervised AI tools to support early uveal melanoma triage and reduce diagnostic variability.

Read paper
0 comments

No comments yet. Be the first.

Related papers

LatestFoundational
AI / Informatics
0·962 for OCT
MerMED-FM was very accurate at diagnosing diseases using eye scans
AI / Informatics
0.98
Artificial intelligence's eyelid-height measurements matched doctors' manual measurements almost perfectly
AI / Informatics
92.1%
combining eye scans and photos correctly told benign from cancerous lesions apart nearly every time
Cohort Study
28.6%
recurred locally in more than 1 in 4 patients over years of follow-up
Cohort Study
-0.395
excision group's post-op eyelid fullness score was lower - greater improvement
Cohort Study
86.7%
of infants probed after 12 months still had unresolved tear duct blockage
Observational
16%
eyelid tissue in rosacea patients showed less of this key repair-signaling protein inside cell nuclei
Cohort Study
25%
about 1 in 4 treated cases had symptoms return after improving