Atrous-based CBAM-Unet framework for intracranial hemorrhage segmentation and volume calculation with grading using CT images

Intracranial hemorrhage (ICH) is a leading cause of death following trauma, often resulting from traumatic brain injury (TBI) when not promptly evaluated and treated within the first 24 hours. Quantitative assessments—such as hemorrhage thickness, volume, and grading—are essential for prognosis but are frequently unavailable. This study aims to develop a reliable segmentation method for ICH to enable accurate volumetric measurement and grading.

We propose an Atrous-based Convolution Block Attention Module (CBAM) Unet for precise feature extraction and robust segmentation of hemorrhagic lesions. The model enhances feature representation through atrous convolutions and attention mechanisms. Hemorrhage volume was subsequently calculated using a pixel-based method. Performance was evaluated using segmentation metrics, classification accuracy, and intraclass correlation coefficient (ICC) for volume measurement reliability.

The proposed method achieved a Dice coefficient of 91.67 and an IoU of 84.62 for hemorrhage segmentation, along with 100% classification accuracy. Volumetric assessment for ICH grading demonstrated excellent reliability, with an ICC of 0.997 and a 95% confidence interval ranging from 0.98 to 1.00, indicating a very high upper bound and strong dependability.

The Atrous-based CBAM-Unet effectively segments intracranial hemorrhages and provides highly reliable volumetric measurements. Its strong performance in both segmentation accuracy and volume estimation supports its potential for clinical application in ICH grading and prognostic evaluation.