Information Geometric Approaches for Patient-Specific Test-Time Adaptation of Deep Learning Models for Semantic...

Area of Science:

• Medical Imaging
• Deep Learning
• Computational Anatomy

Background:

• Existing test-time adaptation (TTA) methods for medical imaging semantic segmentation are often impractical due to constraints, reliance on prior information, or additional networks.
• These limitations can lead to performance deterioration and reduced generalizability in patient-specific applications.

Purpose of the Study:

• To propose a novel, generic, and off-the-shelf framework for regularized patient-specific TTA of deep learning models.
• To address the limitations of current TTA methods by leveraging information geometric principles.

Main Methods:

• Developed a framework treating pre-trained and adapted models as statistical neuromanifolds.
• Applied constrained functional regularization using information geometric measures for TTA.
• Evaluated the approach on diverse medical imaging segmentation tasks.

Main Results:

• Demonstrated improved generalization for segmenting COVID-19 anomalies in CT images.
• Achieved effective cross-institutional brain tumor segmentation from MR images.
• Showcased successful segmentation of retinal layers in OCT images.
• Validated robust patient-specific adaptation without significant computational burden.

Conclusions:

• The proposed information geometry-based TTA framework offers a practical and effective solution for medical image segmentation.
• This novel approach enhances model performance and generalizability across various patient-specific tasks.
• It represents a significant advancement in TTA for medical deep learning applications.