Do you need sharpened details? Asking MMDC-Net: Multi-layer multi-scale dilated convolution network for retinal...

Area of Science:

• Medical Imaging
• Computer Vision
• Deep Learning

Background:

• Convolutional neural networks (CNNs), particularly U-shaped architectures, have advanced retinal vessel segmentation.
• Existing methods often fail to fully utilize global information in fundus images and struggle with the class imbalance between background and blood vessels.

Purpose of the Study:

• To develop a novel deep learning model, MMDC-Net, to improve retinal vessel segmentation by addressing limitations in global information capture and class imbalance.
• To enhance the accuracy and detail of blood vessel identification in fundus images.

Main Methods:

• Designed a Multi-Layer Multi-Scale Dilated Convolution Network (MMDC-Net) based on the U-Net architecture.
• Incorporated an MMDC module for capturing global information across diverse receptive fields via a cascaded approach.
• Introduced a Multi-Layer Fusion (MLF) module to integrate complementary features and filter noise post-decoder.
• Utilized a recall loss function to mitigate the class imbalance issue.

Main Results:

• MMDC-Net demonstrated superior performance in qualitative and quantitative evaluations across multiple fundus image datasets (STARE, CHASEDB1, DRIVE, HRF).
• The model achieved satisfactory accuracy and sensitivity, sharpening key blood vessel details.
• Experiments validated the effectiveness and generalization capabilities of MMDC-Net on datasets with varying resolutions.

Conclusions:

• MMDC-Net effectively captures global information and handles class imbalance, leading to improved retinal vessel segmentation.
• The proposed MMDC and MLF modules enhance feature representation and detail preservation.
• MMDC-Net offers a robust and generalizable solution for accurate retinal vessel segmentation in medical imaging.