On the Detection of Colorectal Polyps with Hierarchical Fine-Tuning

Colorectal cancer is the largest cause of cancer deaths for both men and women, hence early detection of polyps plays an important role for survival. Despite the advancement of the state-of-the-art in the context of objects detection, methods suffer in recognising small polyps, which are the most important to detect since are the one appearing at the beginning of the lesion progression. To deal with the detection of small polyps, we propose to reorganise the training procedure of an object detection in a hierarchical way such that it can better consider visual colonoscopy images content and size polyp variability. Due to the small size of the datasets publicly available in this context, and considering that current datasets do not have an adequate variability to properly assess the performances of a polyp detector, for this study we have collected a new large scale dataset which have been labelled by colonoscopy experts during a clinical trial. The employed dataset is an order of magnitude larger than the datasets currently available in literature and it is better suited to perform more appropriate benchmark of polyps detection since it has been acquired with different colonoscopy devices and because contain a high number of different polyps with real variability in appearance and size. Experimental results on this novel large dataset point out that, by considering the polyps size variability in a hierarchical fine-tuning, polyps can be detected with very high per-frame and per-polyp F1 score of 78.56% and 95.60% respectively.