The challenge focuses on automatically localizing and segmenting colon polyps using a benchmarking multi-center dataset provided by the event organizers. This year, the challenge drew over 400 participants from all over the world. The challenge consisted of two separate tasks, one for polyp localization using bounding boxes and the other for polyp segmentation. 

– We participated in the polyp segmentation task, where we proposed two unique deep neural network architectures named TriUnet and DivergentNets, says PhD student Vajira Thambawita. 

TriUnet is an architecture composed of three U-Nets and was the winning model for the first round of the challenge. 

DivergentNets is an ensemble-based model made up of multiple popular segmentation models, specifically TriUnet, UNet++, FPN, DeepLabv3, and DeepLabv3++, and was the overall winner of the competition.

The code used for this challenge is made publicly available on GitHub. 

The paper titled “DivergentNets: Medical Image Segmentation by Network Ensemble” will be available soon for the complete details of the implementation.