IPM Take
The sharp lesson is that AI screening can fail for ordinary operational reasons. If retinal images are not gradable, the algorithm cannot save the pathway. The Brazilian study matters because it shows that image capture, pupil dilation, camera workflow and local calibration are not technical details. They are access conditions. For diabetic eye care, AI implementation should be judged not only by model performance, but by whether screening can produce usable images in real clinics.
Executive Summary
A 2026 study in Arquivos Brasileiros de Oftalmologia evaluated how mydriasis affects retinal image gradability and AI-based diabetic retinopathy detection using a handheld camera in real-world, resource-limited settings. Mydriatic images were more often gradable than non-mydriatic images, 82.1 percent versus 55.6 percent. The AI model also performed better with mydriatic images, with accuracy of 85.15 percent compared with 79.68 percent in non-mydriatic images. The authors conclude that mydriasis improves image quality and AI performance, while noting that in low and middle income settings, where pharmacologic dilation may be impractical, model calibration and thresholding for non-mydriatic images are essential.
Why it matters
- Clinicians: Need to know when non-dilated images are reliable and when poor image quality should trigger repeat imaging, dilation or referral
- Diagnostics / pathology: Must treat image capture, gradability and quality assurance as core parts of AI screening, not as minor workflow details.
- Public authorities: Should avoid scaling AI eye screening without protocols for staff training, dilation, image quality monitoring and referral.
Before AI retinal screening can expand, health systems need to solve a basic question: can clinics consistently capture images good enough for automated interpretation?
What changed in this study is the focus on the screening environment, not only the AI model. The study compared mydriatic and non-mydriatic images, assessed factors associated with gradability, and externally validated a ResNet-200d model on both image types. Lower gradability in non-mydriatic images was associated with systemic hypertension, older age, male sex and longer diabetes duration.
The affected population is people with diabetes who need diabetic retinopathy screening, especially in settings where specialist access is limited and portable cameras may be used. The implementation challenge is that dilation can improve image quality, but it may also add time, staffing needs, patient discomfort and workflow burden in low-resource settings.
For IPM, the implication is direct. AI eye screening should not be implemented as software alone. It needs trained staff, imaging protocols, quality monitoring, referral pathways and local thresholds that reflect how screening actually happens. Without that, AI may look scalable in theory while failing at the first step of the pathway.
