AECNN: Adversarial and Enhanced Convolutional Neural Networks

Saeed Izadi, Ghassan Hamarneh

Abstract: The proposed method for segmenting gastrointestinal polyps from colonoscopy images uses an adversarial and enhanced convolutional neural networks (AECNN). As the number of training images is small, the core of AECNN relies on fine-tuning an existing deep CNN model (ResNet152). AECNN’s enhanced convolutions incorporate both dense upsampling, which learns to upsample the low-resolution feature maps into pixel-level segmentation masks, as well as hybrid dilation, which improves the dilated convolution by using different dilation rates for different layers. AECNN further boosts the performance of its segmenter by incorporating a discriminator competing with the segmenter, where both are trained through a generative adversarial network formulation.