Adaptively learning local shape statistics for prostate segmentation in ultrasound

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Computational Anatomy

Background:

• Automatic prostate segmentation from 2-D transrectal ultrasound (TRUS) is crucial for clinical applications.
• Existing methods struggle with prostate base and apex segmentation due to significant shape variations.

Purpose of the Study:

• To develop a novel TRUS video segmentation algorithm addressing challenges in prostate base and apex segmentation.
• To improve the robustness and accuracy of prostate gland segmentation in TRUS videos.

Main Methods:

• A new TRUS video segmentation algorithm employing global and patient-specific local shape statistics as constraints.
• Adaptive learning of shape statistics within local neighborhoods to capture patient-specific variations.
• Integration of learned shape statistics into a deformable model for segmentation.

Main Results:

• The proposed method effectively captures patient-specific shape statistics and adapts to local shape changes.
• Robust segmentation of the entire prostate gland, with significantly improved performance in base and apex regions.
• Achieved an average mean absolute distance error of 1.65 ± 0.47 mm across 19 video sequences.

Conclusions:

• The novel algorithm demonstrates superior performance in segmenting the prostate, particularly in challenging regions.
• The use of adaptive, patient-specific shape statistics enhances segmentation accuracy and robustness.
• This method offers a significant advancement for automated prostate segmentation in clinical settings.