Spatial attention-based CSR-Unet framework for subdural and epidural hemorrhage segmentation and classification using...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurosurgery

Background:

• Accurate segmentation of intracranial hemorrhage (ICH) in CT scans is crucial for timely neurosurgical intervention.
• Manual segmentation is labor-intensive and prone to errors, necessitating automated solutions.
• Distinguishing between subdural (SDH) and epidural hemorrhages (EDH) presents a significant challenge in automated detection.

Purpose of the Study:

• To develop and evaluate a deep learning approach for automated detection and segmentation of cerebral bleeding in head CT scans.
• To specifically differentiate between subdural (SDH) and epidural hemorrhages (EDH) with high precision.
• To enhance segmentation efficiency and feature extraction for improved diagnostic accuracy.

Main Methods:

• Utilized a deep learning framework incorporating a Spatial attention-based CSR (convolution-SE-residual) Unet architecture.
• Employed head CT scans as the imaging modality for analysis.
• Focused on rich segmentation and precise feature extraction for enhanced performance.

Main Results:

• The proposed CSR-based Spatial network demonstrated superior performance compared to other models.
• Achieved a mean Dice coefficient of 0.970 and mean Intersection over Union (IoU) of 0.718.
• Specific Dice scores for EDH and SDH were 0.983 and 0.969, respectively, indicating high segmentation accuracy.

Conclusions:

• The CSR Spatial network effectively segments brain hemorrhages, achieving excellent Dice coefficient scores.
• This deep learning model shows potential for improving the meticulousness of fatality prediction in neurological conditions.
• The approach enhances representation learning for complex segmentations, outperforming alternative deep learning techniques.