Hierarchical Pictorial Structures for Simultaneously Localizing Multiple Organs in Volumetric Pre-Scan CT

Area of Science:

• Medical Imaging
• Computer Vision
• Anatomy

Background:

• Simultaneous organ parsing in volumetric CT is crucial for applications like personalized scan planning.
• Clinical pre-scan data varies in dose and quality, posing challenges for organ localization.

Purpose of the Study:

• To propose a novel learning-based framework, hierarchical pictorial structures (HPS), for accurate and efficient organ localization in CT scans.
• To address the challenges of diverse pre-scan data quality and enable applications like personalized radiation dose reduction.

Main Methods:

• Developed a hierarchical model mirroring anatomical decomposition, learning local appearance/shape and probabilistic structural arrangements.
• Integrated pictorial structures into a hierarchical framework for reduced interpretation time and enhanced geometric constraints.
• Utilized probabilistic cost maps from random decision forests with 3D HOG features for fast training and application, ensuring invariance to artifacts.

Main Results:

• Achieved accurate localization of 10 or more salient organs in volumetric CT scans.
• The HPS framework demonstrated high accuracy suitable for clinical applications.
• Processing time for all steps was approximately 3 minutes per scan.

Conclusions:

• The HPS framework offers an efficient and accurate solution for multi-organ segmentation in CT imaging.
• This method facilitates personalized scan planning and dose reporting, particularly with low-dose pre-scan data.
• The approach shows robustness across varied patient demographics and pathologies.