Effective Diabetic Retinopathy Classification With Siamese Neural Network: A Strategy for Small Dataset Challenges

Early detection of diabetic retinopathy, a complication of vision loss in advanced stages of diabetes, is essential to avoid permanent vision impairment. However, the automatic detection of diabetic retinopathy through medical image processing requires a large number of training data to build a model with good performance. This poses a challenge when working with small datasets as these models need large datasets to perform well on unseen data. In this paper, we design a few-shot Siamese Neural Networks combined with pre-trained models, such as VGG16, ResNet50, and DenseNet121, to effectively differentiate between classes using small lesions in the retinal images. The proposed model is trained based on the similarity between the pair of images using a comparatively small dataset and performs well for a five-class classification problem. We use the Fine-Grained Annotated Diabetic Retinopathy (FGADR) and APTOS 2019 Vision Impairment Detection dataset, where a small ratio of training images is used to train the model. To evaluate our model, we conduct the testing on the remaining data and achieve good accuracy when trained on limited images, with fewer epochs and fewer parameters. The proposed model achieves high accuracy rates on five-class classification of 80% on FGADR and 81% on APTOS 2019 datasets, with a consistent quadratic weighted kappa (QWK) score of 0.89 across both datasets. Furthermore, we conduct an in-depth analysis of hyperparameter optimisation, specifically investigating different pair selection techniques, loss functions, and distance layers to thoroughly evaluate their impact on the performance of the model. Our proposed model demonstrates promising results when combined with an attention mechanism to perform multiclass classification of diabetic retinopathy using a limited number of eye fundus images, outperforming existing approaches with only a small number of epochs in training.