XGBoosted Binary CNNs for Multi-Class Classification of Colorectal Polyp Size

Abstract

Colorectal cancer (CRC) is marked by the development of tumors/outgrowths known as polyps. AI-assisted endoscopy is inevitable in the modern world for better and more efficient polyp detection and classification. Often, the risk associated with CRC is indicated by the polyp’s size. Automated size classification of colorectal polyps from endoscopic images is a boon to endoscopists to monitor and diagnose the polyps. While previous research efforts have predominantly centered around the pathological categorization of polyps, limited attention has been directed towards the classification of polyp size. In this paper, we have proposed a deep learning-based model for the multi-class classification of colorectal polyps into four classes: 0–5 mm, 5–10 mm, 10–14 mm, and $>=14$ mm. A narrow range in polyp size classification provides more information about the growth of the polyp as opposed to binary classification. We also show that the One vs Rest classification technique using binary classifiers outperforms the usual approach of using a single CNN for multi-class classification. Also, we use XGBoost with the binary classifiers to further increase the performance of the model. The experimental results report the effectiveness of our proposed model in performing multi-class polyp size classification. The approach is expected to assist clinicians in estimating polyp size efficiently.

Keywords

• Endoscopy
• colorectal polyps
• size classification
• depth maps
• convolutional neural networks
• xgboost