An Improved Deep Neural Network For Nuclei Semantic Segmentation

Convolutional neural networks(CNNs)show outstanding performance in many image processing problems.Semantic segmentation based on deep learning is a pixel-level image processing method,and deep nerual networks can start from the original pixel-wise features,through the layer of convolution operation,and finally extract the high level features which are crucial to the image segmentation,and then use these features automatically divide the object area from the image and identify the category in the object.Since the classical FCN or U-Net often misidentifies the blurred targets in the process of cell nuclei segmentation,in this work we address the task of nuclei semantic segmentation with an improved U-Net and make three main contributions that are experimentally shown to have substantial practical merits.First,we combine batch normalization and a Re LU function after each convolutional layer to solve gradient-vanishing problem.Second,we further propose a hybrid loss function to enhance the overlap between the prediction and the ground truth and reduce the over-segmentation in a non-nuclei image.Third,we introduce a residual path,instead of combining the encoder features with the decoder features in a straight-forward manner,we pass the encoder features through a sequence of convolutional layers to make them have the same depth,these additional non-linear operations are expected to reduce the semantic gap between encoder and decoder features.Our proposed approach has achieved new state-of-the-art results 93% Dice coefficient and90% Jaccard similarity coefficient on validation set which not only surpasses the original U-Net on a data set of 670 images and approximately 29500 cell nuclei published in the Data Science Bowl in 2018,but also significantly better than the classic fully convolutional neural network model FCN-8s.