Polyp Detection in Colonoscopy Images using Improved Deformable DETR

Early detection of colorectal cancer (CRC) can lead to a better prognosis and clinical management. This can be achieved by detecting polyps, the precursor to such cancer. Therefore, in this work, we have proposed a deep learning based polyp detection approach using Deformable DETR. We introduced a loss function, replacing the conventional loss function used in the aforementioned model. We further experimented with the different layers in the model's architecture and reported the optimal number of layers for our work. Exhaustive experiments and analyses are done with different loss functions, augmentations, and TTA (Test Time Augmentation). The results show improvement over the state-of-the-art methods and provide more precise and compact bounding boxes that are close to the clinically validated ground truth.