An Investigation for Colorectal Cancer Early Diagnosis Using Hessian Vector-based Texture Features

Early screening and diagnosis for polyps will significantly reduce their cancerous likelihood since most colorectal cancers are stemmed from neoplastic polyps at their earlier stage. This paper attempted to investigate the accuracy rate of diagnosis on small and medium polyps. To mitigate the bias of this study, a two-step-two-user scheme was adopted to simulate the real scenario in our practice. Two users are required to anticipate the investigation for two different missions, i.e. training and prediction, where Hessian vector is employed to generate texture features. To further study the effect of different growing stages, three training schemes are designed to carry out our investigations, i.e. small-polyp-mode, medium-polyp-mode and mixed-polyp-mode. Experimental outcomes demonstrate that the medium-polyp-mode could bring the best results while the worst results were produced over the small-polyp-mode. By considering the small sizes of polyps in their earlier stage, the results of medium-polyp-mode and the mixed-polyp-mode are acceptable.