A lightweight hierarchical convolution network for brain tumor segmentation.

Area of Science:

• Medical imaging analysis
• Deep learning for neuro-oncology

Background:

• Brain tumor segmentation is crucial for treatment and surgical planning.
• Current 3D convolutional networks are computationally expensive and struggle with multi-tumor segmentation.
• There's a need for efficient models applicable to clinical settings.

Purpose of the Study:

• To propose a lightweight hierarchical convolution network (LHC-Net) for brain tumor segmentation.
• To reduce computational costs while maintaining high segmentation accuracy.
• To improve the segmentation of multiple tumors with uncertain boundaries.

Main Methods:

• Developed LHC-Net, a novel network utilizing hierarchical convolutions with residual-like connections.
• Implemented a multi-scale strategy on channels to enhance feature extraction.
• Replaced common 3D convolutions with hierarchical convolutions to reduce parameters and computation.

Main Results:

• LHC-Net achieved Dice scores of 76.38% (ET), 90.01% (WT), and 83.32% (TC) on the BraTS2020 dataset.
• Outperformed 3D U-Net, achieving better results especially on multi-tumor segmentation.
• LHC-Net has 1.65M parameters and 35.58G FLOPs, significantly less than 3D U-Net.

Conclusions:

• LHC-Net enables automatic segmentation of tumor sub-regions from multi-modal brain MRI.
• The model offers competitive performance with reduced computational demands, suitable for limited medical resources.
• LHC-Net's multi-scale strategy effectively segments multiple brain tumors, showing potential for other segmentation tasks.