Spectral clustering of shape and probability prior models for automatic prostate segmentation.

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Biomedical Engineering

Background:

• Transrectal Ultrasound (TRUS) imaging presents challenges for prostate segmentation due to artifacts and anatomical variability.
• Accurate prostate segmentation is crucial for various clinical applications, including treatment planning and monitoring.

Purpose of the Study:

• To develop an improved computer-aided method for automatic or semi-automatic prostate segmentation in TRUS images.
• To address limitations of traditional statistical models by incorporating posterior probability information.

Main Methods:

• Proposed a novel segmentation approach utilizing multiple mean parametric models derived from principal component analysis (PCA).
• Employed posterior probability maps from random forest classification to build, initialize, and propagate segmentation models.
• Integrated spectral clustering of combined shape and appearance parameters to generate multiple mean models.

Main Results:

• Achieved a high mean Dice Similarity Coefficient (DSC) of 0.96±0.01, indicating excellent segmentation accuracy.
• Demonstrated a rapid mean segmentation time of 0.67±0.02 seconds.
• Validated the method on 46 images from 23 datasets using a leave-one-patient-out framework.

Conclusions:

• The proposed method effectively segments the prostate in TRUS images, outperforming traditional approaches.
• The integration of posterior probability and PCA-derived models enhances segmentation accuracy and efficiency.
• This technique holds promise for improving computer-aided diagnosis and treatment planning in prostate cancer management.