Multi-scale characterizations of colon polyps via computed tomographic colonography

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Texture Analysis

Background:

• Gray-level co-occurrence matrix (GLCM) is a key texture descriptor in medical imaging for computer-aided diagnosis.
• Traditional GLCM methods face challenges like direction sparsity and dense sampling.

Purpose of the Study:

• To enhance GLCM texture features by incorporating multi-scale analysis.
• To introduce a new parameter, 'stride', for GLCM definition.
• To propose and evaluate three multi-scaling GLCM models.

Main Methods:

• Introduced a new 'stride' parameter for GLCM.
• Developed three multi-scaling GLCM models: multiple displacements, learning model by multiple strides (LMS), and multiple angles.
• Tested models on a dataset of 63 computed tomography colonoscopy polyp masses (32 adenocarcinomas, 31 adenomas).
• Compared proposed models against traditional GLCM descriptors and a deep learning model.

Main Results:

• The learning model by multiple strides (LMS) demonstrated superior performance.
• LMS achieved an area under the receiver operating characteristics curve (AUC) of 0.9450 with a standard deviation of 0.0285.
• This represents a significant improvement in prediction power compared to other methods.

Conclusions:

• Multi-scale analysis, particularly the LMS model, substantially enhances GLCM-based texture analysis for medical imaging.
• The proposed LMS model offers improved classification accuracy for polyp masses in computed tomography colonoscopy.
• This approach shows promise for advancing computer-aided diagnosis in medical imaging.