Automated brain structures segmentation from PET/CT images based on landmark-constrained dual-modality atlas...

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Radiology

Background:

• Automated segmentation of brain structures in PET images aids in diagnosing and monitoring brain diseases.
• Existing atlas-based methods often struggle with accuracy for small, deep brain structures due to global optimization strategies.

Purpose of the Study:

• To develop a novel PET/CT-based algorithm for accurate brain volume of interest (VOI) segmentation.
• To improve the precision of atlas registration for deep brain structures using local landmarks and dual-modality information.

Main Methods:

• Combines anatomical atlases with local deep brain landmarks detected via Deep Q-Network (DQN).
• Integrates dual-modality PET/CT information to refine extracerebral contour registration.
• Constrains atlas registration using detected local landmarks for improved accuracy.

Main Results:

• Achieved high accuracy in brain VOI delineation on 86 clinical PET/CT images.
• Reported an average Dice similarity score of 0.79 and an average surface distance of 0.97 mm.
• Demonstrated a volume recovery coefficient close to 1, indicating precise volume measurement.

Conclusions:

• The proposed algorithm effectively optimizes global brain matching and local structure alignment.
• It offers a fully automated solution for high-quality brain structure parcellation from PET/CT images.
• This method enhances the accuracy of automated brain segmentation for clinical applications.