MM-GLCM-CNN: A multi-scale and multi-level based GLCM-CNN for polyp classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Distinguishing malignant from benign lesions is crucial for effective cancer management and early detection.
• Convolutional Neural Networks (CNNs) show promise in medical imaging but require large datasets, which are often unavailable for pathological ground truth.
• Small, pathologically-proven datasets pose a challenge for training accurate CNN models for lesion diagnosis.

Purpose of the Study:

• To develop a CNN model capable of learning features from small, pathologically-proven datasets for differentiating malignant from benign colon polyps.
• To enhance feature extraction by incorporating multi-scale and multi-level texture analysis.
• To improve the diagnostic performance of CNNs in lesion classification using limited data.

Main Methods:

• Proposed a Multi-scale and Multi-level based Gray-level Co-occurrence Matrix CNN (MM-GLCM-CNN) model.
• Utilized Gray-level Co-occurrence Matrix (GLCM) to characterize lesion heterogeneity and texture.
• Implemented an adaptive multi-input CNN framework with an Adaptive Weight Network for feature fusion and refinement.

Main Results:

• The MM-GLCM-CNN achieved an Area Under the ROC Curve (AUC) of 93.99% on small colon polyp datasets.
• Demonstrated a performance gain of 1.49% over existing state-of-the-art methods on the same dataset.
• Validated the effectiveness of incorporating lesion characteristic heterogeneity for malignancy prediction with limited data.

Conclusions:

• The MM-GLCM-CNN model successfully differentiates malignant from benign polyps using texture features from small datasets.
• The approach highlights the significance of lesion heterogeneity in improving diagnostic accuracy for limited pathological data.
• This method offers a viable solution for CNN-based lesion diagnosis when large annotated datasets are unavailable.