HMRNet: High and Multi-Resolution Network With Bidirectional Feature Calibration for Brain Structure Segmentation in...

Area of Science:

• Medical imaging
• Artificial intelligence in oncology
• Radiotherapy planning

Background:

• Accurate segmentation of Anatomical brain Barriers to Cancer spread (ABCs) is vital for defining Clinical Target Volumes (CTVs) in brain tumor radiotherapy.
• Current U-Net variants struggle with the diverse shapes and sizes of ABCs, particularly thin structures like the falx cerebri.

Purpose of the Study:

• To develop an advanced segmentation network, HMRNet, capable of accurately delineating brain tumor radiotherapy targets.
• To improve the segmentation of challenging, thin anatomical structures within the brain.

Main Methods:

• Proposed a High and Multi-Resolution Network (HMRNet) with parallel multi-scale and high-resolution feature learning branches.
• Introduced a Bidirectional Feature Calibration (BFC) block for mutual spatial attention between network branches.
• Implemented a two-stage strategy involving rough localization followed by fine segmentation.

Main Results:

• The two-stage segmentation approach significantly outperformed single-stage methods.
• HMRNet effectively preserved high-resolution information, enhancing segmentation of thin structures.
• The BFC block demonstrated superior performance compared to unidirectional attention mechanisms.

Conclusions:

• The HMRNet with BFC offers a robust solution for accurate ABCs segmentation in radiotherapy.
• This approach achieved second place in the MICCAI 2020 ABCs challenge, showing potential for improved brain tumor CTV delineation.