CHSNet: Automatic lesion segmentation network guided by CT image features for acute cerebral hemorrhage

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Stroke, a cerebrovascular disease, causes significant mortality and disability.
• Accurate segmentation of acute cerebral hemorrhage in CT images is crucial for diagnosis and treatment.
• Existing methods may struggle with the high density, multi-scale, and variable locations of hemorrhage lesions.

Purpose of the Study:

• To propose an automatic segmentation network (CHSNet) for segmenting acute cerebral hemorrhage lesions in cranial CT images.
• To achieve accurate 3D visualization and localization of cranial lesions post-segmentation.
• To improve the clinical adjuvant diagnosis of stroke.

Main Methods:

• Developed a convolutional neural network (CHSNet) with an encoding-decoding backbone, Res-RCL module, Atrous Spatial Pyramid Pooling, and Attention Gate.
• Utilized a dataset of 5998 cranial CT slices from 203 patients with acute cerebral hemorrhage.
• Conducted comparative and ablation experiments to validate model performance.

Main Results:

• CHSNet achieved high segmentation accuracy on two test sets (Test 1: Dice=0.918, IoU=0.853; Test 2: Dice=0.716, IoU=0.604).
• The model successfully enabled 3D visualization and localization of hemorrhage lesions.
• Experimental results demonstrated the effectiveness of the proposed network.

Conclusions:

• CHSNet effectively segments acute cerebral hemorrhage lesions in cranial CT images.
• The 3D visualization and localization capabilities offer significant potential for clinical adjuvant diagnosis in stroke patients.
• The study highlights the utility of AI in improving stroke imaging analysis.