A machine learning pipeline for internal anatomical landmark embedding based on a patient surface model

Area of Science:

• Medical Imaging
• Radiology
• Computational Anatomy

Background:

• Siemens Healthineers introduced patient surface models for diagnostic imaging.
• These models can automate workflows, manage dose, and aid navigation.
• Patient surface models have potential for simplifying interventional imaging.

Purpose of the Study:

• To extend previous patient surface models by incorporating internal anatomical landmarks.
• To investigate methods for accurately positioning these internal landmarks.
• To enable automated positioning of robotic angiography systems for interventional procedures.

Main Methods:

• Trained an active shape model using an avatar database and segmented CT surfaces.
• Implemented a gravity correction procedure for patient models.
• Derived spatial relationships between the patient surface model and internal anatomical landmarks using CT datasets.

Main Results:

• Evaluated the method using 20 datasets with 50 internal landmarks each.
• Compared four generalized linear models for landmark positioning.
• Lasso regression achieved the best mean estimation error over all landmarks.

Conclusions:

• The achieved accuracy is sufficient for automatic positioning of interventional X-ray imaging systems.
• This advancement can facilitate faster and more precise interventional procedures.