Boundary-Weighted Domain Adaptive Neural Network for Prostate MR Image Segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate prostate segmentation in 3D MR images is crucial for cancer diagnosis and treatment.
• Challenges include unclear boundaries, complex textures, and variations in prostate anatomy.
• Deep learning methods, particularly CNNs, require large annotated datasets, which are often scarce in medical imaging.

Purpose of the Study:

• To develop an automated prostate segmentation method that overcomes challenges in MR image analysis.
• To enhance sensitivity to prostate boundaries for more precise segmentation.
• To address the issue of limited annotated data for training deep learning models.

Main Methods:

• Proposed a boundary-weighted domain adaptive neural network (BOWDA-Net).
• Introduced a boundary-weighted segmentation loss to improve edge detection.
• Implemented an advanced boundary-weighted transfer learning approach for small datasets.

Main Results:

• The BOWDA-Net demonstrated increased sensitivity to object boundaries.
• The model achieved superior performance compared to state-of-the-art methods on three diverse MR prostate datasets.
• Boundary-weighted loss and transfer learning effectively addressed segmentation difficulties and data limitations.

Conclusions:

• BOWDA-Net offers a robust solution for automated prostate segmentation in MR images.
• The proposed boundary-aware techniques enhance segmentation accuracy, particularly in challenging cases.
• This approach shows significant promise for improving prostate cancer diagnosis and treatment planning.