An Encoder-decoder Deep Learning Model Combining Mixed Attention
Mechanism and Asymmetric Convolution for Automation of Retinal Vessels
Segmentation
Abstract
The segmentation of the retinal vascular tree is the fundamental step
for diagnosing ophthalmological diseases and cardiovascular diseases.
Most existing vessel segmentation methods based on deep learning give
the learned features equal importance. Ignored the highly imbalanced
ratio between background and vessels (the majority of vessel pixels
belong to the background), the learned features would be dominantly
guided by background, and relatively little influence comes from
vessels, often leading to low model sensitivity and prediction accuracy.
The reduction of model size is also a challenge. We propose a mixed
attention mechanism and asymmetric convolution encoder-decoder
structure(MAAC) for segmentation in Retinal Vessels to solve these
problems. In MAAC, the mixed attention is designed to emphasize the
valid features and suppress the invalid features. It not only identifies
information that helps retinal vessels recognition but also locates the
position of the vessel. All square convolutions are replaced by
asymmetric convolutions because it is more robust to rotational
distortions and small convolutions are more suitable for extracting
vessel features (based on the thin characteristics of vessels). The
employment of asymmetric convolution reduces model parameters and
improve the recognition of thin vessel. The experiments on public
datasets DRIVE, STARE, and CHASE\_DB1 demonstrated that
the proposed MAAC could more accurately segment vessels with a global
AUC of 98.17$\%$, 98.67$\%$, and
98.53$\%$, respectively. The mixed attention proposed
in this study can be applied to other deep learning models for
performance improvement without changing the network architectures.