Intracerebral hemorrhage CT scan image segmentation with HarDNet based transformer.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• U-Net models for hemorrhage segmentation have limitations in parameter efficiency, model size, and speed.
• Existing encoder-decoder architectures struggle with efficient information transfer.

Purpose of the Study:

• To propose TransHarDNet, an improved image segmentation model for diagnosing intracerebral hemorrhage from CT scans.
• To enhance efficiency and inference speed while maintaining high performance in hemorrhage segmentation.

Main Methods:

• Integrating HarDNet blocks into the U-Net architecture.
• Connecting the encoder and decoder using a transformer block.
• Training and testing on a large dataset of 82,636 CT scan images with five hemorrhage types.

Main Results:

• TransHarDNet achieved a Dice coefficient of 0.712 and an IoU of 0.597 on a test set of 1200 hemorrhage images.
• Demonstrated superior performance compared to U-Net, U-Net++, SegNet, PSPNet, and HarDNet.
• Achieved an inference speed of 30.78 frames per second (FPS), faster than most encoder-decoder models.

Conclusions:

• TransHarDNet offers reduced network complexity and improved inference speed for brain hemorrhage segmentation.
• The proposed model maintains high performance, outperforming conventional segmentation models.
• TransHarDNet presents a viable and efficient solution for the automated diagnosis of intracerebral hemorrhage.