Exploring fusion model for the detection and localization of colorectal polyps

Computer Vision (CV) and Artificial Intelligence (AI) have reported unprecedented growth in recent years and demonstrated potential in almost all domains. In the medical realm machine learning and deep learning techniques are widely used for detection as well as classification o fa nomalies and segmentation of anomaly regions. Salient feature extraction is a challenging task in traditional machine learning techniques but deep learning models successfully extract better features with convolutional neural networks (CNN) and report better performance when the training dataset is tremendous. In the medical domain, the main challenge is the limited availability of annotated data to perform training. Colorectal polyp detection and localization is an arduous task due to the close resemblance of polyps with the surrounding colon regions and the irregular shapes of the polyps. But detection, localization, and segmentation of colon polyps are of great importance in the screening and removal of polyp regions for the prevention of colorectal cancer. Currently used U-Net and Mask-RCNN-based architectures for polyp segmentation are computationally expensive. In this paper, we explore a fusion model to overcome the downsides of both machine learning and deep learning techniques. The fusion model serves the similar purpose of segmentation with less computational overhead and can be used to assist clinicians in screening tests. In the proposed work, pre-trained deep CNNs are explored to extract salient features from image patches followed by ensemble machine learning techniques for the detection of colon polyps. A localization algorithm is applied on the detected image patches for localizing the position of the colon polyps.