GLCM-CNN: Gray Level Co-occurrence Matrix based CNN Model for Polyp Diagnosis

The accurate identification of malignant polyp on colon CT is critical for the early detection of colorectal cancer, which also offers patients the best chance of cure. Deep learning based methods, especially convolution neural network (CNN) based methods, have been proposed for computer-aided polyp diagnosis due to CNN's strength in feature learning. However, most of the current CNN models focus on the 2D information or use multiple 2D slices as a 2.5D model input, which does not consider the $3D$ spatial information. In this work, we propose a CNN based 3D polyp diagnosis method. The proposed method encodes the $3D$ information into a multi-dimensional gray-level co-occurrence tensor. Each dimension represents one sampling view in the 3D space and 13 dimensions are used in this work. This model takes advantage of the co-occurrence matrix which is a good texture indicator to differentiate the tissue textures between benign and malignant. Additionally, our proposed method solves the problem of input size selection due to huge variants of polyp size and could be extended to other applications. Experiment results demonstrated that our method achieves an AUC of 0.93, which outperforms 2D (AUC 0.57) and 3D (AUC 0.72) convolution neural network solutions and the current state-of-the-art method (AUC 0.86).