Feature attention improves the performance of a transfer learning-based model in detecting diabetic retinopathy

Abstract

Purpose : Early detection of retinopathy is necessary to prevent vision impairment in patients with diabetes. Previous studies have reported machine learning methods for diabetic retinopathy (DR) classification from retinal images with variable results. We tested the hypothesis that addition of a vessel-segmented image as an attention map to the retinal image improves the performance of an automatic transfer learning-based deep learning model for DR classification.

Methods : Retinal images were captured using a commercially available instrument (Optos Inc., US), from 58 diabetic and 12 non-diabetic subjects. Images were cropped, centered on the optic nerve head. A pretrained double U-net neural network was used to generate a vessel-segmented image. Adding this image to the original retinal image as an additional image channel created the vessel-attention image. A VGG16 model pretrained on ImageNet was used to benefit from transfer learning. Following data augmentation (rotate, zoom, shift, mirror), images were randomly split into 75/25% training/testing datasets. After the initial training, fine tuning was performed. Classification performance with and without the vessel-segmented image was compared.

Results : Seventy images were included from diabetic and non-diabetic subjects (age: 54 ± 15, and 60 ± 12, respectively). Subjects were clinically evaluated and categorized as non-diabetic (N=12), diabetic without DR (N=25), with non-proliferative DR (N=28), or proliferative DR (N=5). Fig 1 illustrates the method used to create the vessel-attention image. The classification accuracy of the model using the original retinal images without adding the vessel-segmented image was 78.2%. The addition of the vessel-segmented image increased the classification accuracy to 92.3%, an 18% increase in performance.

Conclusions : Our preliminary results demonstrate that the addition of a vessel-segmented image to the retinal image has the potential to improve the performance of a deep learning model for detecting diabetic retinopathy in retinal images. Further studies with larger datasets and in other domains are necessary to validate, and evaluate the scalability of our results.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.

 

Vessel-attention Image generation: The retinal image was segmented using a pretrained double U-net network. The resulting vessel-segmented image was added to the retinal image as an additional channel creating the vessel-attention image.