Segmenting CT prostate images using population and patient-specific statistics for radiotherapy

Area of Science:

• Medical Imaging
• Radiotherapy
• Computational Anatomy

Background:

• Accurate segmentation of the prostate in CT images is vital for image-guided radiotherapy.
• Traditional methods struggle to capture intrapatient variations, especially with limited patient data.
• Intrapatient variation is more critical than interpatient variation in treatment-time CT scans.

Purpose of the Study:

• To develop a novel deformable model for segmenting sequential CT prostate images.
• To accurately capture intrapatient variations during radiotherapy using population and patient-specific statistics.
• To improve the robustness and accuracy of prostate segmentation in challenging clinical scenarios.

Main Methods:

• Utilized a weighted combination of gradient and probability distribution function (PDF) features for the appearance model.
• Dynamically adjusted feature weights to optimize model deformation.
• Incorporated an online learning mechanism to build adaptive shape and appearance statistics.
• Calculated optimal gradient profile length to mitigate inconsistencies from adjacent bone and gas regions.

Main Results:

• The gradient-PDF combined features increased the segmentation success ratio by 5.2% (94.1% to 99.3%) compared to traditional gradient features.
• The full online update strategy improved the mean Dice Similarity Coefficient (DSC) from 86.6% to 89.3% (2.8% gain).
• The best performance, achieved with manual modification before online update, yielded a mean DSC of 92.4% and a mean Absolute Surface Distance (ASD) of 1.47 mm.

Conclusions:

• The proposed method achieves accurate and robust prostate segmentation on CT images, even with limited patient samples.
• The novel approach effectively addresses intrapatient variation crucial for radiotherapy.
• The method demonstrates significant potential for clinical application in image-guided radiotherapy.