Non-Euclidean classification of medically imaged objects via s-reps

Area of Science:

• Medical Imaging
• Computational Anatomy
• Machine Learning

Background:

• Accurate classification of medical images is crucial for diagnosis.
• Existing methods often rely on global volume, potentially missing subtle shape differences.
• Non-Euclidean geometric properties offer rich shape information but require specialized handling.

Purpose of the Study:

• To develop a novel classification scheme using skeletal representations for enhanced medical image analysis.
• To evaluate the discriminative power of non-Euclidean geometric properties (GOPs) derived from skeletal representations.
• To assess the impact of Euclideanization on classification performance compared to traditional methods.

Main Methods:

• Proposed a statistical classification method combining distance weighted discrimination (DWD) with Euclideanization of non-Euclidean GOPs from skeletal representations.
• Evaluated the method on 3D hippocampi classification between schizophrenic and healthy subjects.
• Compared performance against global volume-based classification and conventional boundary point distribution models (PDMs).

Main Results:

• The proposed skeletal representation with Euclideanization significantly outperformed global volume-based classification.
• Euclideanized skeletal representations showed superior discriminative power compared to non-Euclideanized versions and PDMs.
• Euclideanized PDMs also demonstrated improved performance over non-linear PDMs, highlighting the benefit of proper Euclideanization.

Conclusions:

• Skeletal representations capture rich, non-Euclidean GOPs beneficial for medical image classification.
• Euclideanization of non-Euclidean GOPs is critical for maximizing classification accuracy.
• The proposed method offers a powerful approach for shape-based medical image analysis, outperforming conventional techniques.