Rotationally resliced 3D prostate TRUS segmentation using convex optimization with shape priors.

Area of Science:

• Medical imaging
• Computational anatomy
• Image processing

Background:

• Accurate prostate boundary segmentation in 3D transrectal ultrasound (TRUS) images is crucial for effective prostate biopsy planning.
• Challenges in segmentation arise from image artifacts and speckle, hindering precise and efficient boundary extraction.

Purpose of the Study:

• To develop an efficient and accurate method for segmenting prostate boundaries from 3D TRUS images.
• To address the limitations posed by poor image quality in existing segmentation techniques.

Main Methods:

• A novel convex optimization-based approach is proposed, reducing 3D segmentation to 2D subproblems.
• A convex relaxation-based contour evolution method incorporates learned 2D nonlinear statistical shape priors for slice-wise segmentation.
• The approach utilizes GPU implementation for high computational performance.

Main Results:

• The method achieved a mean Dice similarity coefficient of 93.4% ± 2.2% in 20 seconds per 3D image.
• Demonstrated superior accuracy and efficiency compared to local-optimization-based methods like level-set and active-contour.
• Exhibited good reproducibility through inter- and intraobserver variability experiments.

Conclusions:

• A semiautomatic segmentation approach for prostate boundary extraction from 3D TRUS images has been developed and validated.
• The proposed method shows promise for clinical application in image-guided prostate biopsy procedures.