Introduction To validate the diagnostic performance of the Eyerobo FC, a new portable non-mydriatic fundus camera for diabetic retinopathy (DR) screening, against an established desktop fundus camera benchmark using a transfer-learning approach in which artificial intelligence (AI)-based detection algorithms trained on desktop images were applied to Eyerobo FC images. Methods This prospective validation study employed a three-tier experimental design. Tier 1 involved training a deep learning model (EfficientNet-B4) on standard desktop camera images from EyePACS and APTOS 2019 datasets. Tier 2 established the reference standard by evaluating the trained model on the Messidor-2 dataset (N = 1748 eyes) captured with a Topcon TRC NW6 desktop camera (sensitivity 92.7%, 95% CI 91.2–94.2%; AUC 0.952, 95% CI 0.943–0.961). Tier 3 validated the same AI model (without retraining) on images from the Eyerobo FC in a prospective cohort (N = 104 eyes: 52 referable DR, 52 non-referable). The primary outcome was noninferiority of sensitivity and specificity (margin <InlineEquation ID="IEq1"> <EquationSource Format="TEX">$\Delta$</EquationSource> <EquationSource Format="MATHML"><math> <mi mathvariant="normal">Δ</mi> </math></EquationSource> </InlineEquation> = 10%) compared with the desktop benchmark. Statistical analysis included bootstrap resampling (1000 iterations) for confidence intervals and a one-sided Z-test for the difference of proportions to assess noninferiority. Results In a balanced cohort of 104 eyes (52 referable DR, 52 non-referable), the Eyerobo FC achieved sensitivity of 92.3% (95% CI 84.4–98.2%) and specificity of 94.2% (95% CI 87.0–100%), demonstrating noninferior performance compared with the desktop benchmark (sensitivity 92.7%, specificity 94.3%). The sensitivity difference of <InlineEquation ID="IEq2"> <EquationSource Format="TEX">$-0.4$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mo>-</mo> <mn>0.4</mn> </mrow> </math></EquationSource> </InlineEquation> percentage points and the specificity difference of <InlineEquation ID="IEq3"> <EquationSource Format="TEX">$-0.1$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mo>-</mo> <mn>0.1</mn> </mrow> </math></EquationSource> </InlineEquation> percentage points were both within the noninferiority margin. AUC was 0.977 (95% CI 0.945–0.997) versus 0.952 for the desktop benchmark. The AI model correctly classified 97 of 104 eyes (93.3% accuracy, 95% CI 88.5–98.1%), with 4 false negatives and 3 false positives. Noninferiority was statistically confirmed for both sensitivity and specificity (P < 0.05). Inter-grader agreement was excellent (Cohen’s kappa = 0.917). Nonmydriatic image gradability rate was 94.4%. Grad-CAM visualization confirmed appropriate model attention to hemorrhages, exudates, and microaneurysms rather than artifacts. Conclusions The Eyerobo fundus camera demonstrates noninferior diagnostic performance (sensitivity 92.3%, specificity 94.2%, AUC 0.977) compared with desktop systems when evaluated with AI algorithms trained exclusively on desktop images. These findings support deploying portable AI-assisted screening in resource-constrained and point-of-care settings, with successful cross-domain transfer learning enabling algorithmic generalizability across imaging platforms.

Validation of the Eyerobo FC Portable Fundus Camera for Diabetic Retinopathy Screening Using Public Datasets and Deep Learning

期刊论文
出版模式：

开放获取
发表日期：

2026年3月11日

查看全文

Emmanuel Eric Pazo,
Xiangying Liu,
Shoukuan Liu,
Qinru Zhang,
Mingxuan Zhang,
Zhihui Zhang,
Salissou Moutari,
Muhammad Usama,
Xinjun Ren

摘要

Introduction

To validate the diagnostic performance of the Eyerobo FC, a new portable non-mydriatic fundus camera for diabetic retinopathy (DR) screening, against an established desktop fundus camera benchmark using a transfer-learning approach in which artificial intelligence (AI)-based detection algorithms trained on desktop images were applied to Eyerobo FC images.

Methods

This prospective validation study employed a three-tier experimental design. Tier 1 involved training a deep learning model (EfficientNet-B4) on standard desktop camera images from EyePACS and APTOS 2019 datasets. Tier 2 established the reference standard by evaluating the trained model on the Messidor-2 dataset (N = 1748 eyes) captured with a Topcon TRC NW6 desktop camera (sensitivity 92.7%, 95% CI 91.2–94.2%; AUC 0.952, 95% CI 0.943–0.961). Tier 3 validated the same AI model (without retraining) on images from the Eyerobo FC in a prospective cohort (N = 104 eyes: 52 referable DR, 52 non-referable). The primary outcome was noninferiority of sensitivity and specificity (margin $\Delta$ $Δ$ = 10%) compared with the desktop benchmark. Statistical analysis included bootstrap resampling (1000 iterations) for confidence intervals and a one-sided Z-test for the difference of proportions to assess noninferiority.

Results

In a balanced cohort of 104 eyes (52 referable DR, 52 non-referable), the Eyerobo FC achieved sensitivity of 92.3% (95% CI 84.4–98.2%) and specificity of 94.2% (95% CI 87.0–100%), demonstrating noninferior performance compared with the desktop benchmark (sensitivity 92.7%, specificity 94.3%). The sensitivity difference of $-0.4$ $- 0.4$ percentage points and the specificity difference of $-0.1$ $- 0.1$ percentage points were both within the noninferiority margin. AUC was 0.977 (95% CI 0.945–0.997) versus 0.952 for the desktop benchmark. The AI model correctly classified 97 of 104 eyes (93.3% accuracy, 95% CI 88.5–98.1%), with 4 false negatives and 3 false positives. Noninferiority was statistically confirmed for both sensitivity and specificity (P < 0.05). Inter-grader agreement was excellent (Cohen’s kappa = 0.917). Nonmydriatic image gradability rate was 94.4%. Grad-CAM visualization confirmed appropriate model attention to hemorrhages, exudates, and microaneurysms rather than artifacts.

Conclusions

The Eyerobo fundus camera demonstrates noninferior diagnostic performance (sensitivity 92.3%, specificity 94.2%, AUC 0.977) compared with desktop systems when evaluated with AI algorithms trained exclusively on desktop images. These findings support deploying portable AI-assisted screening in resource-constrained and point-of-care settings, with successful cross-domain transfer learning enabling algorithmic generalizability across imaging platforms.