Landmark Localization From Medical Images With Generative Distribution Prior

Area of Science:

• Medical Image Analysis
• Machine Learning
• Computer Vision

Background:

• Anatomical landmarks in medical images possess inherent structural information.
• Accurate localization of these landmarks is crucial for various diagnostic and analytical tasks.
• Existing methods may face limitations in effectively leveraging prior structural knowledge.

Purpose of the Study:

• To improve medical landmark localization by incorporating a learned distribution prior.
• To introduce a novel framework, Normalizing Flow-based Distribution Prior (NFDP), for enhanced localization.
• To optimize the integration of distribution priors into regression-based localization models.

Main Methods:

• Modeling landmark distribution using normalizing flows.
• Integrating a flow-based landmark distribution prior as a learnable objective function.
• Employing an integral operation for differentiable heatmap-to-coordinate mapping.

Main Results:

• NFDP demonstrated high-fidelity outputs across three X-ray-based landmark localization datasets.
• The method achieved superior prediction accuracy compared to existing techniques.
• The normalizing flows module was detached during inference, minimizing computational burden.

Conclusions:

• NFDP offers an efficient and effective approach for medical landmark localization.
• The proposed method achieves a strong balance between prediction accuracy and inference speed.
• NFDP provides a valuable tool for advancing medical image analysis applications.