Major Temporal Arcade Segmentation through Hyperparameter Fine-Tuning in an Attention-Based U-Net

Abstract

This study presents an innovative approach
based on a U-Net architecture for the automatic
segmentation of the major temporal arcade (MTA) in
fundus images. The strategy consists of attention
and segmentation procedures. Vascular structure
segmentation using the popular VGG-16 deep learning
model and transfer learning using the U-Net architecture
constitute the initial phase. An attention module is added
in the second stage to capture the MTA shape and
remove any vascular structures other than MTA. The
effectiveness of the suggested method is evaluated in
terms of F1 score, sensitivity, specificity, and accuracy,
and compared with various state-of-the-art vascular
segmentation methods. A data augmentation approach
was applied only to the training set. An expert categorized
MTA images from the popular DRIVE database, which
included 40 fundus images, to determine the vascular
structure. 1377 training patches were obtained by
dividing each training image into 81 patches and applying
data augmentation to ensure a reliable training procedure.
A comparative study was conducted, showing how the
best metrics were achieved, as well as comparing them
with other methods for segmenting arterial structures.
The best results were obtained with an accuracy of
0.9923 and an F1 of 0.7700 using the test set of fundus
images. The numerical results of the suggested method
on the automatic MTA segmentation problem outperform
those of the comparative approaches.