A Point Says a Lot: An Interactive Segmentation Method for MR Prostate via One-Point Labeling

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Biomedical Engineering

Background:

• Accurate prostate segmentation in magnetic resonance (MR) imaging is crucial for diagnosis and treatment planning.
• Existing segmentation methods often require extensive manual input or struggle with precise boundary delineation.

Purpose of the Study:

• To develop and evaluate a novel MR prostate segmentation method utilizing minimal user input (one-point labeling).
• To investigate the enhancement of segmentation performance through the integration of boundary detection and patch-based prediction.

Main Methods:

• A novel segmentation approach was developed, incorporating physician-provided single-point labeling within the prostate region.
• Location-prior maps were generated based on proximity to the clicked point and boundary information.
• A multi-channel fully convolutional network processed MR images and prior maps for patch-based prediction.
• A level-set method was employed for final prostate segmentation using the predicted likelihood map.

Main Results:

• The proposed method demonstrated promising performance in MR prostate segmentation on a dataset of 22 patient MR images.
• One-point labeling significantly enhanced segmentation results compared to pure patch-based prediction methods.
• The integration of boundary detection and patch-based prediction proved effective in improving accuracy.

Conclusions:

• The developed one-point labeling method offers an efficient and effective approach for MR prostate segmentation.
• This technique has the potential to streamline the segmentation process and improve clinical workflow.
• Further validation on larger datasets is warranted to confirm generalizability.