Fuzzy U-Net Neural Network Design for Image Segmentation

Abstract

One of the problems in biomedical image analysis is the problem of nuclei segmentation. The annotation of nuclei by hand has proven itself very time consuming with varying results depending on many factors. More recently convolutional neural networks have made the problem of automatic image segmentation easier, faster and more reliable. In this article, an extension to the standard U-Net is proposed which aims at improving the quality of biomedical image segmentation. By integrating Fuzzy computations in the standard U-Net architecture we have achieved even better accuracies than the ones reached by the base architecture. The Fuzzy Layers resemble a sense of uncertainty, which is already seen in the real world. This allows for a more precise detection and instance segmentation of cellular nuclei. When the original U-Net was first developed one of its focuses was to be trainable with a small dataset. This quality has been proved useful when undertaking the task of Biomedical Image Segmentation. The model was trained with the Kaggle 2018 Dataset. The segmentation process comes right after, including the following two steps: 1) creating a prediction matrix, 2) thresholding the matrix to attain a visual result. The second step exhausts the following techniques: Manual Thresholding; Adaptive Thresholding; Gaussian Thresholding, and Otsu Thresholding. The results point out which thresholding technique, combined with a certain set of Fuzzy Layers, yields the best possible results in terms of accuracy of the models.