Quantum-Enhanced EfficientNet-B0 for Multi-Class Retinal Disease Classification on Fundus Images

Abstract Vision loss caused by retinal diseases, Diabetic Retinopathy, Glaucoma, and Retinitis Pigmentosa is a growing concern worldwide. While deep learning has made strides in medical imaging, many current models are still limited to just a few classification categories, which isn't ideal in clinical practice. In this study, we tackled this issue by building a model that can classify ten different retinal conditions from fundus images. We used EfficientNet-B0 as our starting point and also developed a hybrid version by inserting a 6-qubit, 6-depth quantum circuit into the network. We trained both models on the Eye Disease Image Dataset. with original dataset, the classical model reached 75.61% validation accuracy and 81.62% on training, while the quantum-enhanced version did slightly better with 77.45% and 83.07%, respectively. After applying data augmentation, performance improved a lot 86.37% validation for the classical model and 93.86% for the quantum one. The hybrid model especially helped in tricky cases like CSCR and Macular Scar. These results demonstrate the potential of quantum-classical integration in the both dataset for medical image classification.