JMIR Publications

ABSTRACT

Background:

Diabetic retinopathy (DR) grading becomes more difficult when models are applied to images from new clinical sources. Differences in cameras, image quality, and patient populations can reduce model performance after transfer. Hybrid quantum-classical models can reduce the number of parameters that need to be trained, but previous quantum studies for DR have been generally evaluated on a single dataset. The key question is whether hybrid quantum-classical learning can be useful for cross-domain DR grading performance across new datasets while reducing task-specific training from millions of parameters to thousands.

Objective:

We developed a parameter-efficient hybrid quantum-classical model for DR grading and evaluated how well it transfers to external datasets. We compared it with fully fine-tuned classification and regression models and examined overall performance, grade-level errors, and confidence calibration.

Methods:

We trained all models on APTOS 2019 and tested them without further tuning on MESSIDOR-2 (n=1,744), IDRiD (n=516), and UoA (n=200). Our Hybrid-QC model uses a frozen ConvNeXt-Tiny encoder followed by a small classical projection, an 8-qubit variational quantum circuit, and a final classification layer. Hybrid-QC updates 25,093 parameters, compared with about 28.05 million trainable parameters in the fully fine-tuned ConvNeXt-Tiny classifier (ConvNeXt-Cls) and regression model (ConvNeXt-Reg), approximately 1,118 times fewer. We first evaluated no-DR versus advanced-DR classification and then the more difficult five-class grading.

Results:

For no-DR versus advanced-DR classification, Hybrid-QC achieved an area under the receiver operating characteristic curve (AUROC) of 0.986 on MESSIDOR-2, 0.989 on IDRiD, and 0.995 on UoA. All three models achieved AUROC above 0.98 on every external dataset, and no two models differed by more than 0.01. In five-class grading, Hybrid-QC achieved quadratic weighted kappa (QWK) values close to those of ConvNeXt-Cls on MESSIDOR-2 (0.608 vs 0.618) and IDRiD (0.719 vs 0.717). ConvNeXt-Reg achieved the highest QWK and the lowest mean absolute error on all three external datasets. Hybrid-QC achieved the highest Grade-1 F1 on MESSIDOR-2, IDRiD, and UoA (0.186, 0.162, and 0.400) and predicted fewer Grade-1 cases as Grade 0 than ConvNeXt-Cls. It also had a lower expected calibration error than ConvNeXt-Cls on MESSIDOR-2 (0.047 vs 0.296), IDRiD (0.141 vs 0.373), and UoA (0.100 vs 0.216), and remained better calibrated even after temperature scaling of ConvNeXt-Cls.

Conclusions:

Hybrid-QC reduces the trainable parameter requirement from about 28 million to about 25 thousand while achieving binary AUROC within 0.01 of the fully fine-tuned models on all three external datasets and five-class QWK close to ConvNeXt-Cls on MESSIDOR-2 and IDRiD. ConvNeXt-Reg remained stronger for overall ordinal grading, while Hybrid-QC achieved the highest Grade-1 F1 on all three external datasets and lower calibration error than ConvNeXt-Cls after domain shift. These results support Hybrid-QC as a parameter-efficient approach to cross-domain DR grading.