Automated Colorectal Polyp Analysis Using EfficientNet-B0, Vision Transformer, and Attention U-Net

Abstract

Even after attempts being made to curb the disease, colorectal cancer remains among the main reasons for deaths from cancer worldwide. The cause of colorectal cancer can be attributed to the development of polyps in the colon and rectum. Early detection and accurate classification of polyps are important in the identification and management of the disease. Conventional methods used to detect polyps have been associated with inadequate lesion detection, quality differences in images, complex tissue structure, and need for clinical expertise. Small lesions with low contrast and concealment increase the difficulty in the whole process. As a way of addressing these challenges, there is an attempt to develop an intelligent approach that automatically detects and classifies polyps. In developing the approach, the images that will be used are those of ColoVision Polyp Dataset (CVPD). With a technique that uses efficient feature representation, contextual understanding, and attention-based localization, it becomes possible to ensure significant areas detection and classify the colon polyps effectively. Detection, segmentation, and classification done simultaneously ensure a complete solution without decreasing efficiency of the process. Several metrics like accuracy, precision, recall, F1 score, DSC, IoU, and mAP are used for analyzing the performance of the system. From the experiment, it can be understood that the proposed solution works well in detecting the polyps as well as classifying the different kinds of polyps such as adenomas, hyperplastic and malignant polyps.