An Adaptive Boosting Strategy for GLCM-CNN Model in Differentiating the Malignant from Benign Polyps

Recently, deep learning such as Convolutional Neural Network (CNN) has shown its superior for the task of image classification. However, it faces great challenges in the medical imaging field, especially for tumor classification in computer-aided diagnosis, because there are many uncertainties of lesions including their size, scaling factor, rotation, shapes, etc. Hence, texture pattern is an option to be fed into the CNN model and among the texture patterns, gray level co-occurrence matrix (GLCM) can be chosen as the texture pattern for its several good properties such as uniform size, shape invariance, posture robustness, scaling invariance. For a 3D volume sample, instead of studying on the raw images, 13 independent GLCMs could be extracted based on the 13 digital distinct directions, and different sampling displacements have been recorded accordingly. Therefore, considering this multi-scale sampling is hypothesized to be optimal compared to without considering the multi-scale nature, we proposed an adaptive boosting learning (ABL) to examine the proposed multi-scale hypothesis. Following the multi-scale nature, the 13 directions can be grouped as three by different sampling displacements, and the ABL is applied to learn the best classification performance from the 3 subgroups. By comparing our ABL with and without the multi-scale nature, our experiments obtained a gain of 0.5~3.7% in terms of AUC (area under the ROC curve) for polyp classification, eventually reaching AUC of 91.13% for a small database of 63 samples with pathological reports. The gains reveal that the multi-scale GLCM-CNN model with ABL strategy can reduce the challenges for classification of limited number of tumors in medical imaging field.