Localized Statistical Shape Models for Large-Scale Problems With Few Training Data.

Area of Science:

• Biomedical image analysis
• Machine learning
• Computer vision

Background:

• Statistical shape models (SSMs) are crucial for biomedical image analysis, aiding tasks like organ segmentation and deep learning data augmentation.
• However, training SSMs requires large datasets, often unavailable, leading to poor representation of local shape details in unseen data.
• Existing methods struggle with data scarcity, limiting the applicability of SSMs in real-world scenarios.

Purpose of the Study:

• To introduce a novel kernel-based method for statistical shape model localization to address limitations in data availability.
• To enhance the representation of local shape details in unseen data within SSMs.
• To enable the efficient generation of flexible and specific shape models from limited training samples for applications like data augmentation.

Main Methods:

• A kernel-based localization technique is proposed, building upon advances in multi-level shape model localization and Grassmannian-based level fusion.
• The method utilizes distance-based covariance matrix manipulations for efficient computation.
• A normalizing flow-based density estimation approach is introduced to improve model specificity.

Main Results:

• The kernelized formulation demonstrated effectiveness on public datasets (JSRT/SCR chest X-ray, IXI brain).
• The proposed density estimation method significantly improved model specificity.
• The method enables the creation of specific shape models even with few training samples.

Conclusions:

• Combining kernel theory and normalizing flows allows for computationally efficient generation of flexible and specific shape models from limited data.
• The developed method and its open-source implementation facilitate the creation of SSMs for data augmentation and other biomedical imaging applications.
• This approach overcomes data scarcity challenges in building robust statistical shape models.