Multi-scale object equalization learning network for intracerebral hemorrhage region segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accurate segmentation of intracerebral hemorrhage (ICH) in CT images is crucial for pathology analysis.
• Existing methods struggle with the complex morphology and multi-scale nature of ICH regions.

Purpose of the Study:

• To develop a novel deep learning network for precise ICH region segmentation.
• To address limitations in capturing discriminative features across different ICH scales.

Main Methods:

• Proposed MOEL-Net (Multi-scale Object Equalization Learning Network) incorporating SFEM, DFEM, and MSFEFM modules.
• Utilized shallow and deep feature extraction with multi-level semantic feature equalization fusion.
• Evaluated on two newly collected datasets: VMICH and FRICH.

Main Results:

• MOEL-Net achieved high Dice scores: 88.28% (VMICH), 90.92% (FRICH Test A), and 90.95% (FRICH Test B).
• Outperformed fourteen existing segmentation methods.
• Demonstrated robust feature capture for diverse ICH segmentation tasks.

Conclusions:

• MOEL-Net effectively segments ICH regions by leveraging multi-scale information and feature equalization.
• The model shows significant potential for advancing automatic clinical ICH segmentation.
• The developed datasets (VMICH, FRICH) will facilitate further research in this area.