Enhancing perihematomal edema segmentation: integrating prior knowledge with deep learning for enhanced accuracy and...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurosurgery

Background:

• Perihematomal edema (PHE) in spontaneous intracranial hemorrhage (ICH) is a critical factor contributing to increased intracranial pressure and poor patient outcomes.
• Accurate segmentation of PHE is vital for effective clinical management and treatment strategies.
• Current deep learning (DL) segmentation methods face challenges due to the indistinct boundaries and grayscale overlap between PHE and surrounding brain tissues.

Purpose of the Study:

• To enhance the accuracy and reliability of PHE segmentation in spontaneous ICH.
• To develop a novel deep learning model that leverages prior knowledge of PHE for improved segmentation performance.

Main Methods:

• Proposed the Perihematomal Edema Synergistic Enhancement Network (PESE-Net) for PHE segmentation.
• Developed a slice similarity-based method to identify relevant PHE slices within ICH images.
• Implemented a feature weighting strategy to synergistically fuse global change features and spatial information of PHE.

Main Results:

• PESE-Net demonstrated strong performance across multiple evaluation metrics when compared to state-of-the-art methods.
• The proposed method achieved the best performance in terms of relative volume difference, indicating accurate volume estimation.
• Experimental results validated the effectiveness of PESE-Net in segmenting PHE regions.

Conclusions:

• PESE-Net effectively utilizes contextual relationships between consecutive slices as prior knowledge for PHE segmentation.
• The model achieves robust and accurate segmentation of PHE, leading to stable volume estimation.
• This approach holds promise for improving the efficiency and reliability of clinical diagnosis in spontaneous ICH management.