LightEyes: A Lightweight Fundus Segmentation Network for Mobile Edge Computing.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Color fundus image analysis aids in diagnosing various eye diseases.
• Deep learning has advanced fundus image segmentation but often results in complex models with slow inference and high memory usage, hindering mobile deployment.
• Existing deep models struggle with efficiency on mobile edge devices.

Purpose of the Study:

• To design a lightweight fundus segmentation network for efficient deployment on mobile devices.
• To improve segmentation accuracy for tiny fundus structures by leveraging high-resolution representations and local features.
• To address the trade-off between segmentation accuracy and computational cost in deep fundus segmentation models.

Main Methods:

• Proposed a lightweight segmentation model named LightEyes.
• Designed a high-resolution backbone network to preserve spatial relationships in feature maps.
• Employed at most 16 convolutional filters per layer to reduce memory usage and training complexity.
• Validated the model on three segmentation tasks (hard exudate, microaneurysm, vessel) across five public datasets.

Main Results:

• LightEyes demonstrated competitive segmentation accuracy and speed compared to state-of-the-art models.
• Achieved inference speeds of 1.6 images/s on Cambricon-1A and 51.3 images/s on GPU.
• The model utilizes only 36,000 parameters, significantly reducing memory footprint.

Conclusions:

• LightEyes offers an efficient solution for deploying deep fundus segmentation models on mobile devices.
• The proposed architecture effectively balances segmentation performance with computational efficiency.
• This work facilitates wider accessibility of advanced diagnostic tools for eye diseases through mobile technology.