MINet: Multi-scale input network for fundus microvascular segmentation.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Vessel segmentation in fundus images is crucial for diagnosing ophthalmic diseases.
• Current deep learning methods struggle with segmenting thin, microscopic vessels due to information loss in deep networks.

Purpose of the Study:

• To develop a novel deep learning model, MINet, for more accurate fundus vessel segmentation, especially for microscopic vessels.
• To enhance the preservation and utilization of microvessel features during segmentation.

Main Methods:

• Proposed a multi-input network (MINet) incorporating a multi-input fusion (MIF) module in the encoder to capture multi-scale features and preserve microvessel information.
• Introduced a multi-scale atrous spatial pyramid (MASP) module to aggregate multi-scale information without resolution loss.
• Implemented a refinement module to improve the detail recovery of segmentation results.

Main Results:

• MINet achieved a high F1 score of 0.8324 on the microvessel segmentation task.
• Demonstrated superior performance compared to current mainstream models in segmenting challenging microscopic vessels.
• Validated on public datasets (HRF, CHASE_DB1) and a combined dataset including Ultra-widefield (UWF) images, showing good generalization ability.

Conclusions:

• MINet effectively addresses the limitations of existing methods in segmenting microscopic fundus vessels.
• The proposed network architecture enhances diagnostic assistance for ophthalmic diseases through improved vessel segmentation accuracy.