Designed to be inspected, tested, and challenged
The current technology is an inference research package, not a complete clinical product. A partner engagement starts by testing one model, one modality, and one defined device domain.
Partner-device input
Color fundus photographs for DR or glaucoma, or OCT B-scans for AMD/DME. A future study must lock the acquisition and preprocessing contract for the selected device.
Separate research model
A dedicated EfficientNetV2-M classifier processes one task. The three models are evaluated separately; the manuscript does not validate one integrated workflow.
Research output
Class probabilities and a thresholded label are produced for analysis. These outputs are not clinical recommendations and must not drive patient care.
Validation analysis
Evaluate discrimination, fixed-threshold transport, calibration, confidence intervals, gradability, subgroups, and failure cases on the target domain.
The device domain matters
Different optics, field of view, resolution, compression, acquisition protocols, and patient populations can shift model behavior. The manuscript does not establish cross-device product compatibility.
That is precisely where a manufacturer partnership creates value: jointly defining and measuring performance on the partner's real image domain.
Known portability gates
- Thresholds: fixed operating points did not transport uniformly across glaucoma datasets.
- Calibration: probability calibration shifted externally and requires site/device-specific assessment.
- ONNX: the DR export did not meet the prespecified logit-parity tolerance; OCT parity was report-stated rather than independently reproduced.
- Workflow: latency, gradability, OOD handling, human factors, and integration were not clinically validated.
What success looks like
A reproducible, independently analyzed result on a defined device and population - followed by a prospective plan only if the evidence gates pass.