Retinal-imaging and referral-support research for preterm infants
microscopIA is developing an integrated research system for retinal imaging in preterm infants. The device is intended to support a high-resolution camera, primarily a compatible smartphone, to improve inspection and capture of infant retinal images.
Research models analyze de-identified retinal images to study image quality and features of retinopathy of prematurity (ROP). The intended goal is to help identify infants with concerning findings or higher risk for referral to a ROP-trained ophthalmologist, who retains diagnosis and treatment decisions. Future use is being evaluated for appropriately trained neonatologists, neonatal nurses and physicians, including general practitioners where permitted, working under approved institutional protocols and local scope-of-practice rules.
Human-led research pathway
- Existing eligibility and follow-up protocol
- Image acquisition by a trained professional
- Image-quality check and research AI analysis
- Qualified clinical review
- Established specialist referral and follow-up
Ungradable image or failed acquisition → repeat or escalate under the approved protocol. There is no AI-created ‘no referral’ pathway.
Latest supplied AI development record
Task-specific development benchmarks—not clinical performance.
Source last modified 19 November 2025. Values are shown by model task because the supplied record does not establish performance of an integrated device–AI system.
Stage-related image classification
92.03Accuracy · reported scoreGCViT · 1,099 public-dataset imagesPlus-disease classification
93.10Accuracy · reported scoreResNet50 · public development datasetVessel segmentation
76.81%Intersection over UnionU-Net · HVDROPDBMethods, additional scores and limitations
- Stage-related classification
- Precision 93.39; recall/sensitivity 92.03; F1 92.32. Classes were Normal, Stages 1–3 and laser scars. Stages 4 and 5 were absent, so these results cannot support safe referral prioritization or deprioritisation.
- Plus-disease classification
- Precision 88.9; recall/sensitivity 89.8; F1 89.4. Dataset size, patient-level split and confidence intervals were not supplied.
- Segmentation
- IoU 76.81% using HVDROPDB and U-Net. The supplied record does not document external validation.
- Not yet evidenced
- The proposed binary ROP route has no supplied benchmark. Zone classification was not trained because the record reports that no suitable public dataset was identified.
The stage and plus values are reproduced as reported; the source does not print a percent sign for them. These are separate public-dataset development tasks with incomplete evaluation details. They are not directly comparable, externally or clinically validated, or evidence of device-level performance. No specificity, AUROC or confidence intervals are claimed.