LLRHNet: Multiple Lesions Segmentation Using Local-Long Range Features.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep convolutional neural networks (CNNs) excel in medical image segmentation but struggle with long-range feature extraction due to convolutional locality.
• Limitations exist in capturing features across layers and distant image regions using traditional CNNs.

Purpose of the Study:

• To develop a novel network, the Local-Long Range Hybrid Features Network (LLRHNet), for enhanced medical image segmentation.
• To address the limitations of CNNs by integrating local and long-range feature extraction capabilities.

Main Methods:

• LLRHNet utilizes an encoder-decoder architecture with iterative aggregation to fuse local features across layers.
• Transformer technology with multi-head self-attention is employed to extract long-range features from image patches.
• Hybrid features are fused to guide up-sampling operations for precise tissue localization.

Main Results:

• LLRHNet demonstrated state-of-the-art performance on two distinct medical image segmentation datasets.
• Evaluated on a public liver lesion dataset (3DIRCADb) and an in-house stroke dataset (SWMH).

Conclusions:

• The proposed LLRHNet effectively integrates local and long-range features for superior medical image segmentation.
• LLRHNet offers a promising approach for improving the accuracy and robustness of medical image analysis tools.