Spatial Bayesian modeling of GLCM with application to malignant lesion characterization.

Area of Science:

• Radiomics and medical image analysis
• Statistical modeling in oncology
• Machine learning applications in cancer research

Background:

• Cancer radiomics transforms medical images into quantifiable data for tumor analysis.
• Gray-level co-occurrence matrices (GLCM) are used for textural feature extraction but often yield redundant information.
• Current methods for radiomic feature analysis lack robustness and clear predictive power.

Purpose of the Study:

• To present a Bayesian probabilistic modeling framework for GLCM.
• To apply this framework for improved cancer detection using computed tomography.
• To overcome limitations of traditional GLCM feature extraction and analysis.

Main Methods:

• Developed a Bayesian probabilistic model treating GLCM as a multivariate object.
• Utilized latent Gaussian Markov random field structure to capture spatial dependencies.
• Applied the model to computed tomography data for adrenal lesion classification.

Main Results:

• The proposed Bayesian framework achieved 81% accuracy in predicting adrenal lesion pathology.
• Outperformed current practices, which had a maximum accuracy of 59%.
• Demonstrated the utility of multivariate Gaussian spatial processes for GLCM analysis.

Conclusions:

• The Bayesian probabilistic modeling framework offers a more effective approach to cancer radiomics.
• This method enhances diagnostic accuracy by avoiding redundant and correlated features.
• The findings support the application of advanced spatial processes for improved medical image analysis.