Improved Polyp Segmentation of Irregular Shapes in Colonoscopy Images

Being able to identify polyps during colonoscopy is critical for cancer prevention. This research proposes a deep neural network model that combines a dilated convolution module with a criss-cross attention-based network to segment polyps from the endoscopic images of the colon. The model utilizes a base network of ResNet50 with global average pooling for feature extraction and employs a criss-cross attention mechanism to gather the context information of all pixels in an image more efficiently. During training, data augmentation techniques such as rotations, flips, scaling, and contrast were used to further improve the model. We conducted a comprehensive set of experiments to test and evaluate the proposed model. The combination of the dilated convolution module, criss-cross attention, and the global average pooling was found to be effective for the polyp segmentation. Our architecture demonstrates better results across all metrics.