Wasserstein-based texture analysis in radiomic studies.

Area of Science:

• Radiomics and Medical Image Analysis
• Computational Pathology
• Machine Learning in Healthcare

Background:

• Radiomics extracts quantitative features from medical images, but traditional methods often lose spatial information from texture matrices like GLCMs.
• Existing texture features in radiomics can be redundant and fail to capture the full spatial context within images.
• Accurate classification of diseases like COVID-19 from CT scans is crucial for timely diagnosis and treatment.

Purpose of the Study:

• To propose a novel radiomics pipeline that preserves spatial information in texture features for improved image classification.
• To introduce new spatial texture features using the Wasserstein metric to quantify similarity within image classes.
• To evaluate the pipeline's effectiveness in diagnosing COVID-19 using CT images.

Main Methods:

• Developed a pipeline incorporating novel spatial texture features derived from Gray-Level Co-occurrence Matrices (GLCMs).
• Utilized the Wasserstein metric from optimal mass transport theory to measure spatial similarity between texture matrices.
• Employed a Support Vector Machine (SVM) classifier with Bayesian optimization for selecting optimal texture references and classification.

Main Results:

• The proposed pipeline, using optimized spatial texture features with the Wasserstein metric, demonstrated high potential in classifying unseen COVID-19 CT images.
• Compared to traditional statistical features and Euclidean-based methods, the Wasserstein-based spatial features showed superior performance.
• The method effectively captures and leverages spatial information inherent in texture matrices for enhanced classification accuracy.

Conclusions:

• The novel radiomics pipeline offers a promising approach for disease classification by preserving and utilizing spatial texture information.
• The Wasserstein metric provides a robust way to quantify spatial similarity for feature extraction in radiomics.
• The developed method and available code can be adapted for various classification tasks in medical imaging beyond COVID-19 detection.