Fast automated segmentation of multiple objects via spatially weighted shape learning.

Area of Science:

• Medical Imaging
• Computational Anatomy
• Cancer Treatment Planning

Background:

• Active shape models (ASMs) are effective for automatic segmentation using shape and appearance priors.
• Accurate prostate segmentation is crucial for effective prostate cancer treatment planning.
• Traditional ASMs require computationally intensive initialization, often involving 3D image registration.

Purpose of the Study:

• To present a fast, self-initialized active shape model (ASM) approach.
• To enable simultaneous fitting of multiple objects using hierarchical, spatially weighted shape learning.
• To improve upon the computational efficiency of ASM-based segmentation.

Main Methods:

• Developed a self-initialized ASM that bypasses complex registration steps.
• Employed hierarchical control with spatially weighted shape learning for object initialization.
• Validated the method on 3D MR images from 38 prostate cancer patients.

Main Results:

• Achieved Dice's similarity coefficients (mean, median, inter-rater) of (0.79, 0.81, 0.85), comparable to multi-atlas methods.
• Demonstrated no registration error, indicating successful self-initialization.
• Reduced computational time to 12-15 minutes, nearly an order of magnitude faster than multi-atlas approaches.

Conclusions:

• The proposed fast, self-initialized ASM offers a computationally efficient alternative for prostate segmentation.
• This method maintains high segmentation accuracy while significantly decreasing processing time.
• The hierarchical, spatially weighted approach effectively initializes and fits multiple objects.