On the Exploitation of Temporal Redundancy to Improve Polyp Detection in Colonoscopy

Colonoscopy is currently the most effective screening method to find precancerous colon polyps and plan their removal. Computer-aided polyp detection can reduce polyp miss detection rates and help doctors find the most critical regions to pay attention to. The challenge in detecting polyps is due to the polyp's morphology and size, and these fall into false-negative. Indeed, polyps may exhibit high variability in shapes (e.g., depressed, flat, pedunculated, etc ...). Moreover, the water injected from the endoscope results in artifacts which impede the detection, and the lubricating mucus causes light artifacts due its glossiness. To address this problem, we propose a mask-based attention mechanism to ensure that the employed detector focuses on particular regions of the image in order to reduce misdetection rate. Our contribution takes advantage of information on polyp's position over time within a video sequence. We provide such information through a binary mask which points out the last-known polyp's position. The proposed approach is validated on a dataset that has been labeled by colonoscopy experts. It contains about 200 videos and more than 500 different polyps with high variability in size and textures. Experimental results show that the proposed attention mechanism recover a smaller number of false negatives and achieves an Fl-score of 80.21%.