A quantum-inspired cross-attention fusion network for multi-class eye disease classification from retinal fundus images

Automated screening of blinding eye diseases from colour fundus photographs is an active area of medical image analysis, yet many reported systems are validated on a single dataset and offer limited insight into their internal decision process. This study introduces QICAF-Net, a dual-branch deep learning framework that couples a Swin Transformer global-context branch with an EfficientNet local-lesion branch and fuses their representations through a novel Quantum-Inspired Cross-Attention Fusion (QICAF) module. The QICAF module reformulates cross-attention using the mathematical formalism of quantum states: token features are amplitude-encoded as complex unit vectors, transformed by a learnable norm-preserving unitary gate, compared through a fidelity-inspired similarity, normalised by a Born-rule measurement rule, and combined through an entanglement-inspired bilinear interaction. All operations are executed as classical linear algebra on conventional graphics hardware; no physical quantum computer and no quantum speed-up are claimed. The model was trained and evaluated on a balanced four-class retinal dataset of 4,217 fundus images spanning cataract, diabetic retinopathy (DR), glaucoma and normal categories. Under stratified five-fold cross-validation with six-view test-time augmentation (TTA), QICAF-Net attained a macro-F1 of 0.9358 ± 0.0113, an accuracy of 0.9367 ± 0.0112 and a macro one-versus-rest area under the receiver operating characteristic curve (AUC) of 0.9914 ± 0.0023; a three-seed evaluation of the full model gave a macro-F1 of 0.9420 ± 0.0071. An 18-run multi-seed component ablation (six fusion configurations × three seeds) confirmed that the quantum-inspired components achieve accuracy statistically indistinguishable from strong classical fusion baselines, and a paired three-seed significance test between QICAF-Net and a matched dual-branch concatenation model returned no statistically significant accuracy difference (paired t-test p = 0.887; Wilcoxon p = 1.000). Against five single-backbone baselines under an identical preprocessing and TTA protocol, QICAF-Net was competitive but did not exceed the strongest backbones (Swin-Tiny 0.9583, ViT-Small 0.9487). The principal contribution of the module is therefore a novel and interpretable formulation rather than a raw accuracy gain. External validation on a heterogeneous independent cohort (N = 3,907) showed heterogeneous cross-dataset transfer - better preserved for cataract and normal and substantially degraded for glaucoma; under an observed external-domain shift the diabetic-retinopathy sensitivity fell to 39.61%, and under colour-constancy correction with six-view TTA the external accuracy was 0.8741 (95% CI 0.8633-0.8841) and the macro-F1 0.7188 (95% CI 0.6969-0.7398). Two complementary explainability methods together with visualisation of the QICAF attention indicated that the model attends to clinically plausible regions, most strongly the optic nerve head for glaucoma. Calibration, quantitative explanation-faithfulness (deletion/insertion) and profiler-based computational analyses were additionally performed. QICAF-Net is thus presented as an interpretable and transparently evaluated framework whose accuracy is on par with established fusion strategies while contributing a new attention formulation to the retinal imaging literature.