Progressive Multiscale Consistent Network for Multiclass Fundus Lesion Segmentation.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Ophthalmology

Background:

• Multi-class segmentation of fundus lesions is challenging due to variations in lesion scale and shape.
• Existing methods struggle with feature interaction and consistency between different scales.
• Lack of interaction leads to deviation and loss of detail, while feature conflicts decrease prediction accuracy.

Purpose of the Study:

• To propose a novel network for effective multi-scale information integration in fundus lesion segmentation.
• To address the issues of feature deviation and conflict in multi-scale feature learning.
• To improve the accuracy and feature representation for multi-class fundus lesion segmentation.

Main Methods:

• Introduced the progressive multi-scale consistent network (PMCNet).
• Developed the progressive feature fusion (PFF) block to integrate adjacent multi-scale features.
• Incorporated the dynamic attention block (DAB) to manage conflicts in multi-scale features.

Main Results:

• The PFF block progressively integrates fine-grained details and high-level semantics.
• The DAB dynamically learns attentive cues to smooth conflicts in multi-scale features.
• PMCNet demonstrated superior performance compared to state-of-the-art methods on three public datasets.

Conclusions:

• The proposed PFF and DAB blocks effectively address multi-scale and feature inconsistency issues.
• PMCNet enhances feature representation for multi-class fundus lesion segmentation.
• The method offers a significant improvement for automated analysis of retinal images.