HMNet: Hierarchical Multi-Scale Brain Tumor Segmentation Network.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Accurate automatic brain tumor segmentation is crucial for clinical practice.
• Convolutional neural networks (CNNs) show promise for automatic segmentation.
• Existing methods face challenges with computational overhead and feature adaptability.

Purpose of the Study:

• To propose a novel hierarchical multi-scale segmentation network (HMNet) for efficient and accurate brain tumor segmentation.
• To address the computational cost of 3D convolutions in CNNs.
• To improve the adaptability of segmentation networks to various tumor shapes and sizes.

Main Methods:

• Developed a hierarchical multi-scale segmentation network (HMNet) with high-resolution and parallel multi-resolution branches.
• Introduced a lightweight conditional channel weighting block to reduce GPU memory and enhance efficiency.
• Proposed a lightweight multi-resolution feature fusion (LMRF) module to decrease model complexity and feature map redundancy.

Main Results:

• HMNet achieved Dice similarity coefficients of 0.781 (ET), 0.901 (WT), and 0.823 (TC) on the BraTS 2020 dataset.
• Comparative experiments demonstrated satisfactory performance against state-of-the-art (SOTA) approaches on multiple datasets.
• The proposed lightweight modules effectively reduced computational overhead and model complexity.

Conclusions:

• HMNet offers an accurate and efficient solution for automatic brain tumor segmentation.
• The network's architecture and novel modules enable effective handling of diverse tumor characteristics.
• HMNet shows significant potential for clinical application in neuro-oncology.