ADASSM: Adversarial Data Augmentation in Statistical Shape Models From Images.

Area of Science:

• Medical Imaging
• Computational Anatomy
• Deep Learning

Background:

• Statistical Shape Models (SSM) are crucial for analyzing anatomical variations in populations, aiding in pathology detection and treatment planning.
• Traditional SSM methods require extensive preprocessing of medical images (CT/MRI), which is time-intensive.
• Deep learning-based Image-to-SSM networks offer efficiency but suffer from data-hunger and overfitting, often relying on offline shape augmentation.

Purpose of the Study:

• To address the texture bias and data limitations in deep learning Image-to-SSM networks.
• To introduce a novel on-the-fly data augmentation strategy for enhancing Image-to-SSM frameworks.
• To improve the accuracy and robustness of deep learning models for shape analysis in medical imaging.

Main Methods:

• Developed a novel on-the-fly data augmentation strategy using data-dependent noise generation (texture augmentation).
• Trained the augmentation framework adversarially against the Image-to-SSM network.
• Generated diverse and challenging noisy samples to train the network.

Main Results:

• The proposed texture augmentation strategy significantly improved the accuracy of Image-to-SSM networks.
• The model learned to focus more on the underlying anatomical geometry rather than superficial pixel values.
• Achieved comparable or improved accuracy to traditional SSM methods while mitigating deep learning model limitations.

Conclusions:

• On-the-fly texture augmentation is an effective method to enhance deep learning Image-to-SSM networks.
• This approach mitigates texture bias and improves generalization in the presence of limited medical data.
• The findings suggest a promising direction for developing more robust and accurate computational anatomy tools.