A U-shaped CNN with type-2 fuzzy pooling layer and dynamical feature extraction for colorectal polyp applications

Abstract

This study aims to propose type-2 fuzzy pooling in a U-shaped convolutional neural network (CNN) architecture (T2FP_UNet). A CNN consists of convolutional, pooling, a fully connected layer, and activation functions. The pooling layer executes a fuzzy pooling operation, utilizing type-2 fuzzy membership function. In contrast to conventional methods (max and average pooling), the fuzzy pooling operation assigns membership values to pixels before computing fuzzy values, thereby preventing the encoder from losing features. The decoder implements dynamic feature extraction to acquire informative features. This approach improves the robustness and uncertainty handling of semantic image segmentation tasks using a modified U-Net architecture with type-2 fuzzy pooling layer and dynamic feature extraction. This method combines the advantages of the feature-fused U-Net architecture, type-2 fuzzy logic and dynamical feature extraction for handling complex uncertainties in image data. Comparative results are tabulated.