Intracerebral Hemorrhage Segmentation on Noncontrast Computed Tomography Using a Masked Loss Function U-Net Approach

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Intracerebral hemorrhage (ICH) volume is critical for predicting outcomes in acute hemorrhagic stroke.
• Accurate hematoma segmentation is essential for volume estimation and feature extraction (e.g., spot sign, texture, iodine content).
• Manual and semi-automatic segmentation methods are time-consuming, subjective, and require expert personnel.

Purpose of the Study:

• To develop and evaluate a convolutional neural network (CNN) for automated intracerebral hemorrhage (ICH) segmentation on noncontrast computed tomography (CT) scans.
• To provide an efficient and objective alternative to manual segmentation for ICH volume estimation and analysis.

Main Methods:

• A U-Net architecture incorporating a masked loss function was employed.
• The model was trained on down-sampled (256x256) noncontrast CT images.
• Data augmentation and a soft Dice loss function were utilized to optimize performance and prevent overfitting.

Main Results:

• The model achieved a median Dice coefficient of 75.9% and a Hausdorff distance of 2.65 pixels, indicating strong segmentation performance.
• Detection accuracy was quantified with a sensitivity of 77.0% and a specificity of 96.2%.

Conclusions:

• The proposed masked loss U-Net demonstrates accuracy in the automatic segmentation of intracerebral hemorrhage.
• Future work should aim to enhance detection sensitivity and compare this model's performance against alternative architectures.