Computer Vision News - September 2022

43 Polyp2Seg: Improved Polyp Segmentation... and I think the resultsandvalidation helped convince the reviewers. We compared all the models using the same protocols so that the comparison was fair . We also carriedout comprehensive ablation studies. Another appealing aspect was our hybrid approach combining transformers, CNNs, and the medical perspective . ” Javier also points out that they tried to write down every detail that could be relevant for others to reproduce their results – this may be part of what made them a winner and also a teaching tool for other groups on how to make a winning paper themselves. Does Javier have a final piece of advice for his fellow researchers? “ Before you start thinking about blocks and architectures, try to understand the problem by getting some inspiration and prior knowledge from a medical point of view, ” he tells us. “ Be inspired by the literature. Combine techniques. Focus on more than one domain. Ultimately, try to answer the question: what technique will get us there? ” “ I found the scarcity of medical imaging data the most challenging part of this work. The architecture is trained from scratch, but without enough data, it tends to overfit. ” He explains that reducing the number of model parameters and including some prior medical knowledge was vital . They have been able to reduce the parameters but recognize that some more work is needed before the method can be applied to real-time polyp detection and segmentation, which is the long-term goal. However, their hard work has paid off. Winning a Best Paper award is not something that happens every day; they should both be immensely proud. “ I never expected it – I didn’t believe it the first time I heard! ” Vittorino laughs. Javier recalls: “ We knew the work was of good quality because it was selected for an oral presentation, but we weren’t expecting to win the Best Paper award. ” Despite their modesty, Javier offers some insight into why they scooped the top prize: “ We did solid work and solid experiments,