RFE-UNet: Remote Feature Exploration with Local Learning for Medical Image Segmentation.

Area of Science:

• Computer Vision
• Medical Image Analysis
• Deep Learning

Background:

• Convolutional Neural Networks (CNNs) have limited remote modeling ability.
• Transformers excel at global modeling but struggle with local details.
• Medical image segmentation requires both local detail and distance modeling capabilities.

Purpose of the Study:

• To address limitations in CNNs and transformers for medical image segmentation.
• To propose a novel module for improved remote and local feature learning.
• To introduce a new multi-organ segmentation dataset (MOD) for evaluation.

Main Methods:

• Development of a remote feature exploration (RFE) module.
• Integration of RFE module to assist local feature generation using remote elements.
• Creation and utilization of the MOD dataset alongside the Synapse dataset for validation.

Main Results:

• The proposed method achieved 79.77% DSC on the Synapse dataset.
• The method achieved 75.12% DSC on the newly created MOD dataset.
• HD95 scores were 21.75 mm on Synapse and 7.43 mm on MOD, indicating improved segmentation accuracy.

Conclusions:

• The RFE module effectively enhances medical image segmentation by leveraging remote features for local detail generation.
• The proposed approach demonstrates superior performance on both established and novel medical imaging datasets.
• This work offers a promising direction for developing more accurate medical image segmentation networks.