Supervised Image Classification Algorithm Using Representative Spatial Texture Features: Application to COVID-19...

Area of Science:

• Computer Vision
• Medical Imaging Analysis
• Machine Learning for Healthcare

Background:

• Texture analysis is crucial for image perception and data analysis across various fields.
• Identifying representative 'good' and 'bad' samples is key to improving predictive model performance.
• Existing methods lack robust ways to define and utilize representative texture samples.

Approach:

• Proposes novel spatial texture features derived from gray-level co-occurrence matrices (GLCMs).
• Integrates these features into a supervised image classification pipeline using Support Vector Machines (SVM).
• Employs Bayesian optimization and the Wasserstein metric from optimal mass transport (OMT) for selecting optimal GLCM references and defining new features.

Key Points:

• Sample fitness is determined by Wasserstein distance and Spearman rank correlation within classes.
• New texture features are calculated as Wasserstein distances between selected references and other samples.
• The method was evaluated for diagnosing COVID-19 using computed tomographic (CT) images.

Conclusions:

• The proposed pipeline effectively incorporates representative texture features for improved image classification.
• This approach shows promise for enhancing diagnostic accuracy in medical imaging, specifically for COVID-19 detection.
• The integration of OMT and Bayesian optimization offers a novel strategy for feature selection and model enhancement.