3D-GLCM CNN: A 3-Dimensional Gray-Level Co-Occurrence Matrix-Based CNN Model for Polyp Classification via CT...

Area of Science:

• Medical imaging analysis
• Computational pathology
• Machine learning in oncology

Background:

• Accurate colorectal polyp classification is crucial for colorectal cancer (CRC) early detection and treatment.
• Convolution neural networks (CNNs) show promise in image recognition tasks, including medical image analysis.
• Differentiating malignant from benign polyps is challenging, especially with limited data.

Purpose of the Study:

• To explore the efficacy of CNNs utilizing gray-level co-occurrence matrices (GLCMs) for classifying colorectal polyps.
• To compare CNN performance on GLCMs versus raw CT images and Random Forest (RF) on Haralick features.
• To assess the potential of this method in a clinical setting for polyp differentiation.

Main Methods:

• Volumetric CT images of 32 malignant and 31 benign colorectal polyps were used.
• Thirteen GLCMs were computed for each polyp from 13 directions.
• CNNs were trained on GLCMs and multi-slice CT images; RF was used for Haralick features derived from GLCMs.

Main Results:

• CNNs using GLCMs achieved AUC scores of 0.91 (two-fold) and 0.93 (leave-one-out).
• Random Forest on Haralick features yielded AUCs of 0.84/0.86.
• CNNs on multi-slice CT images achieved AUCs of 0.79/0.80.
• CNNs with GLCMs outperformed other methods, especially with a small dataset.

Conclusions:

• CNN learning from GLCMs effectively differentiates malignant from benign colorectal polyps.
• This approach overcomes limitations of small datasets and improves classification performance.
• The method shows significant potential for clinical application in polyp diagnosis.