Flow Matching-Based Data Synthesis for Robust Anatomical Landmark Localization

Area of Science:

• Medical Imaging
• Deep Learning
• Computer Vision

Background:

• Anatomical landmark localization (ALL) is vital for medical imaging applications like therapy planning and surgery.
• Deep learning models for ALL often struggle with small datasets, leading to overfitting and poor generalization.
• Lack of large, annotated medical datasets hinders the development of robust ALL models.

Purpose of the Study:

• To propose a novel generative approach using Flow Matching for synthesizing diverse, annotated medical images for ALL data augmentation.
• To address the challenges of limited data and overfitting in deep learning-based ALL.
• To improve the robustness and generalization capabilities of ALL models.

Main Methods:

• A multi-channel generative approach utilizing Flow Matching to synthesize medical images paired with multi-channel heatmaps encoding landmark configurations.
• Automatic assessment of synthetic image-heatmap pair quality using a Statistical Shape Model for landmark plausibility and Fréchet Inception Distance for image quality.
• Integration of Flow Matching-generated synthetic data into the training process of an ALL network.

Main Results:

• Flow Matching synthesized image-heatmap pairs demonstrated superior quality and diversity compared to Generative Adversarial Networks and diffusion models.
• ALL networks trained with Flow Matching data showed improved robustness, particularly with limited training data and occlusions.
• Synthetic data generated via Flow Matching effectively augmented training datasets for ALL tasks.

Conclusions:

• Flow Matching offers a powerful method for generating high-quality, diverse annotated medical images for ALL data augmentation.
• The proposed approach significantly enhances the robustness and generalization of deep learning models for ALL, outperforming existing generative methods.
• This technique holds promise for advancing medical imaging analysis and interventions by overcoming data scarcity challenges.