Automatic Detection and Segmentation of Brain Hemorrhage Based on Improved U-Net Model.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Brain hemorrhage is a critical cause of mortality, necessitating accurate and early detection of bleeding in the brain.
• Existing methods for cerebral hemorrhage localization often lack the precision required for detailed clinical assessment.
• Accurate segmentation of brain damage is crucial for effective and timely patient treatment.

Purpose of the Study:

• To develop an automated method for precise brain hemorrhage detection and segmentation using enhanced U-Net architectures.
• To evaluate the performance of U-Net models with DenseNet-121, ResNet-50, and MobileNet-V2 backbones for cerebral hemorrhage segmentation.
• To compare the proposed models against previous works for cerebral CT image analysis.

Main Methods:

• Proposed an improved U-Net model by integrating DenseNet-121, ResNet-50, and MobileNet-V2 as feature extraction backbones.
• Utilized a transfer learning approach for training the models on a Kaggle head CT dataset with bleeding and non-bleeding classes.
• Focused on automatic detection and segmentation of brain hemorrhage regions, moving beyond simple bounding box localization.

Main Results:

• Achieved high segmentation accuracy, reaching up to 99% on the head CT dataset.
• Demonstrated the effectiveness of U-Net architecture modifications for medical image segmentation with minimal preprocessing.
• Provided comparative results highlighting the suitability of the proposed models for cerebral CT image analysis.

Conclusions:

• The proposed enhanced U-Net models offer a highly accurate solution for automatic brain hemorrhage detection and segmentation.
• The method is efficient, requiring minimal preprocessing and performing well even with smaller datasets.
• This advancement holds significant potential for improving diagnostic accuracy and patient outcomes in cases of cerebral bleeding.