Bargain Speed for Performance in Colonoscopic Video Sequence Segmentation with ConvNeXt and KAN

Colorectal cancer is one of the most spread types of cancer. Because the performance of the standard screening procedure is greatly influenced by the skills of the endoscopist performing it, there is a high need of improvements. For this purpose we developed a solution based on Neural Networks which can segment colonoscopies videos providing the specialists the zone where polyps reside. The model combines the technique of Neighbourhood Attention from the PNS+ model with the recently proposed Kolmogorov-Arnold Networks in order to better process both long-term and short-term information, obtained from input video sequences, utilizing present as well as past frames. Some changes to the original baseline implementation have been performed, such as introducing the ConvNeXt model, Group Norm, Mish activations and the AdamW optimizer. The developed solution manages to greatly improve the performance of the base model, with a Dice of 0.82 (from 0.73), and compared to more recent models it provides relatively small performance reductions with higher inference speeds of up to 5 times faster.