MetricUNet: Synergistic image- and voxel-level learning for precise prostate segmentation via online sampling

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Fully convolutional networks (FCNs) like UNet and VNet are standard for semantic segmentation.
• Conventional FCN training with cross-entropy or Dice loss can yield non-smooth segmentations, especially in low-contrast CT prostate images.

Purpose of the Study:

• To develop an improved semantic segmentation method for prostate delineation in CT images.
• To address the limitations of conventional FCNs in producing smooth and accurate segmentations.

Main Methods:

• A two-stage framework was proposed: initial prostate region localization followed by precise segmentation.
• A multi-task UNet architecture with a novel online metric learning module using voxel-wise sampling was introduced.
• The network features dual branches for segmentation and voxel-metric learning, trained end-to-end.

Main Results:

• Ablation studies demonstrated that the proposed method learns more representative voxel-level features than conventional loss functions.
• The method achieved superior performance in CT prostate segmentation compared to state-of-the-art approaches.
• The online metric learning module effectively improved segmentation quality and smoothness.

Conclusions:

• The proposed two-stage framework with online voxel-wise metric learning significantly enhances CT prostate segmentation accuracy.
• This novel approach offers a more robust solution for medical image segmentation tasks with low tissue contrast.
• The method provides a promising advancement for automated prostate segmentation in clinical settings.