HF-UNet: Learning Hierarchically Inter-Task Relevance in Multi-Task U-Net for Accurate Prostate Segmentation in CT...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiation Oncology

Background:

• Accurate prostate segmentation is crucial for effective external beam radiation therapy.
• Current multi-task deep learning methods may overlook task-specific feature representations, limiting performance.
• Prostate boundaries in CT images can be unclear, posing segmentation challenges.

Purpose of the Study:

• To develop an advanced deep learning model for precise prostate segmentation in CT images.
• To address the limitations of conventional multi-task learning in medical image segmentation.
• To improve the accuracy and reliability of prostate segmentation for radiation therapy planning.

Main Methods:

• A two-stage deep learning network was proposed, featuring a fast localization stage followed by an accurate segmentation stage.
• A multi-task learning framework was employed, incorporating prostate segmentation and prostate boundary delineation.
• A novel hierarchically-fused U-Net (HF-UNet) architecture with attention-based task consistency learning was introduced to balance shared and task-specific feature learning.

Main Results:

• The proposed HF-UNet demonstrated superior performance compared to conventional multi-task networks.
• Extensive evaluations on planning CT and prostate zonal datasets confirmed the effectiveness of the method.
• HF-UNet achieved state-of-the-art results in prostate segmentation accuracy.

Conclusions:

• The HF-UNet architecture effectively learns shared and task-specific representations for improved prostate segmentation.
• The proposed method offers a significant advancement for precise prostate segmentation in radiation therapy.
• HF-UNet provides a robust and accurate solution for challenging prostate segmentation tasks in CT imaging.