Hierarchical multi-resolution deep encoder-decoder network for MRI Brain Tumor segmentation

Area of Science:

• Computer Vision
• Deep Learning
• Medical Image Analysis

Background:

• Multi-scale encoder-decoder architectures leverage multi-resolution information for image segmentation.
• Existing methods vary in formulation and application of multi-resolution strategies.
• Current architectures often require full-resolution inputs and single-step inference, limiting efficiency and applicability.

Purpose of the Study:

• To categorize and formulate multi-resolution concepts in deep encoder-decoder segmentation networks.
• To address shortcomings of existing models, including high computational/memory demands and inefficient single-step problem-solving.
• To introduce Hierarchical Multi-Resolution deep encoder-decoder Networks (HMRNets) for improved efficiency and performance.

Main Methods:

• Formulated multi-resolution concepts within deep encoder-decoder segmentation networks.
• Introduced Hierarchical Multi-Resolution deep encoder-decoder Networks (HMRNets), trained hierarchically from low to high resolution.
• Designed LAUNet, a lightweight attentive U-shaped network, as a baseline HMRNet.

Main Results:

• HMRNets allow segmentation of lower-resolution images, suitable for memory/hardware constraints.
• The hierarchical learning approach enhances discriminative performance by effectively utilizing multi-resolution information.
• LAUNet, as an HMRNet baseline, achieved competitive performance on brain tumor segmentation datasets (Decathelon, BraTS18, BraTS20).

Conclusions:

• HMRNets offer a viable solution for efficient and effective image segmentation, especially in resource-limited scenarios.
• The hierarchical learning strategy improves segmentation accuracy.
• LAUNet demonstrates the potential of HMRNets for state-of-the-art performance in medical image segmentation.