Covariant Image Representation with Applications to Classification Problems in Medical Imaging

Area of Science:

• Medical Image Analysis
• Computer Vision
• Differential Geometry

Background:

• Traditional image analysis treats images as functions with invariant intensity values under domain transforms.
• This functional paradigm is becoming insufficient with advanced imaging technologies and diverse image types.
• Covariant images, where values jointly vary with domain transforms, offer a new perspective.

Purpose of the Study:

• To formally introduce the concept of covariant images.
• To propose a novel similarity measure for covariant images, incorporating both extrinsic and intrinsic geometries.
• To demonstrate the utility of this similarity measure in medical image classification.

Main Methods:

• Formal definition of covariant images, focusing on symmetric positive-definite tensor fields and Gaussian mixture fields.
• Development of a similarity measure for covariant images viewed as embedded shapes in a product space.
• Application of the similarity measure within a supervised learning framework for classification tasks.

Main Results:

• The proposed similarity measure effectively captures relationships between covariant images.
• Demonstrated success in classifying brain MR images for age and Alzheimer's disease status.
• Achieved effective seizure detection from high angular resolution diffusion magnetic resonance imaging scans.

Conclusions:

• The covariant image framework and proposed similarity measure offer a powerful new approach for medical image analysis.
• This method enhances classification accuracy in challenging medical imaging datasets.
• The approach has potential for broader applications in analyzing complex image data where values covary with domain transforms.