Interpretable exemplar-based shape classification using constrained sparse linear models.

Area of Science:

• Medical imaging analysis
• Computational anatomy
• Disease pattern recognition

Background:

• Diseases often cause visible changes in organ shape.
• Accurate shape classification can aid in disease identification and understanding disease relationships.

Purpose of the Study:

• To develop a holistic framework for shape classification applicable to disease detection.
• To create a method that generalizes to unseen shapes without feature extraction.
• To provide interpretable results by identifying similar shape exemplars.

Main Methods:

• Utilized a lossless scalar field representation for shapes.
• Employed non-parametric classification based on sparse recovery.
• Developed a holistic framework bypassing traditional feature extraction.

Main Results:

• Achieved accurate classification of three cerebellar diseases and controls in patient data.
• Demonstrated robust performance on publicly available 2D datasets, including ETH-80.
• Attained 88.4% classification accuracy on the ETH-80 dataset.

Conclusions:

• The proposed shape classification framework offers accurate and interpretable disease identification.
• The method shows promise for clinical applications in diagnosing organ-related diseases.
• Generalizability to unseen shapes and interpretable outputs enhance diagnostic utility.