Vision Transformer-based recognition of diabetic retinopathy grade

In the domain of natural language processing, Transformers are recognized as state-of-the-art models, which opposing to typical convolutional neural networks (CNNs) do not rely on convolution layers. Instead, Transformers employ multi-head attention mechanisms as the main building block to capture long-range contextual relations between image pixels. Recently, CNNs dominated the deep learning solutions for diabetic retinopathy grade recognition. However, spurred by the advantages of Transformers, we propose a Transformer-based method that is appropriate for recognizing the grade of diabetic retinopathy.

The purposes of this work are to demonstrate that (i) the pure attention mechanism is suitable for diabetic retinopathy grade recognition and (ii) Transformers can replace traditional CNNs for diabetic retinopathy grade recognition.

This paper proposes a Vision Transformer-based method to recognize the grade of diabetic retinopathy. Fundus images are subdivided into non-overlapping patches, which are then converted into sequences by flattening, and undergo a linear and positional embedding process to preserve positional information. Then, the generated sequence is input into several multi-head attention layers to generate the final representation. The first token sequence is input to a softmax classification layer to produce the recognition output in the classification stage.

The dataset for training and testing employs fundus images of different resolutions, subdivided into patches. We challenge our method against current CNNs and extreme learning machines and achieve an appealing performance. Specifically, the suggested deep learning architecture attains an accuracy of 91.4%, specificity = 0.977 (95% confidence interval (CI) (0.951-1)), precision = 0.928 (95% CI (0.852-1)), sensitivity = 0.926 (95% CI (0.863-0.989)), quadratic weighted kappa score = 0.935, and area under curve (AUC) = 0.986.

Our comparative experiments against current methods conclude that our model is competitive and highlight that an attention mechanism based on a Vision Transformer model is promising for the diabetic retinopathy grade recognition task.