Segmentation of kidneys using a new active shape model generation technique based on non-rigid image registration

Area of Science:

• Medical Imaging
• Computer Vision
• Computational Anatomy

Background:

• Active shape models (ASMs) are vital for image segmentation.
• Current methods struggle with complex shapes and require parametric formulations.
• Establishing point correspondences is a key challenge in ASM construction.

Purpose of the Study:

• To develop a novel technique for solving the point correspondence problem in ASMs.
• To improve the detail and efficiency of ASM creation and updating.
• To evaluate the performance of the new method against established algorithms.

Main Methods:

• Utilizing non-rigid image registration to establish point correspondences.
• Applying the method to a training set of 3D kidney pairs from CT images.
• Comparing the registration-based approach with a minimum description length (MDL) formulation.

Main Results:

• The novel method yields more detailed ASMs without explicit parametric formulations.
• Segmentation accuracy improved, with sensitivity increased by approximately 10%.
• Mean squared error was reduced by 20-34%, demonstrating enhanced segmentation precision.

Conclusions:

• Non-rigid image registration offers a superior approach to solving the point correspondence problem for ASMs.
• The proposed method leads to more accurate 3D medical image segmentation.
• The technique is efficient for updating ASMs with new prior knowledge.