Medical Image Segmentation Based On ResDense U-Net

In recent years,with the rapid development of Deep Learning,medical image segmentation technology has made great breakthroughs.In this respect,U-Net has been the most popular architecture in medical image segmentation.U-Net is an end-to-end symmetric model that shows excellent performance in segmenting multimodal medical images,providing convenience for doctors to diagnose and treat patients.However,through many experiments on medical data sets,it is found that the traditional U-Net is inadequate in some aspects.Therefore,this paper proposes a new deep learning model to improve the classic U-Net.The new model is based on the ideas of DenseNet and ResNet and U-Net network structure,so it is called ResDense U-Net.I conducts many tests and comparisons on the skin lesion data set and the optical cell data set.Compared with the classic U-Net,the performance of ResDense U-net is respectively improved by 3.07% and 1.38%.Although the new model structure has not changed much,the performance improvement is still obvious and the details in imaging are more perfect than U-Net.The innovations of this paper are as follows: Firstly,based on the ideas of ResNet and DenseNet,the continuous double-layer convolutions in the original U-Net are replaced by ResDense-blocks to ease the disappearance of gradients and enhance feature reuse.Considering that the number of U-Net original channels has been doubled step by step,the output channels in ResDense-blocks have been adjusted a little to reduce the parameters and prevent overfitting.Secondly,replace all basic convolutions with asymmetric convolution blocks whose convolution kernel skeleton is "ten" shaped.Although the training time is increased,the accuracy is improved.The Layer Normalization(LN)is used to improve the model performance,so the stability of the model no longer depends on the value of batch size.Thirdly,the original UNet was directly stitched with the shallow features extracted from the left half of the network and the deep features extracted from the right half of the network.This direct stitching of deep features and shallow features may have a negative impact on network expression.The new model adds a series of magnitude Res-blocks to the long connections,that is four,three,two,one.The top-level long connection adds four Res-blocks and so on.Given that there is a certain semantic difference between deep and shallow features.The closer to the upper layer,the larger the difference and the closer to the bottom of the "U",the smaller the difference.I have also considered using Dense-block,however,there are more deep and shallow problems.Res-block can better solve deep network problems,so Res-block is the choice.When the medical image is segmented,the region of interest often only accounts for a small part and it may spread throughout the image,which greatly increases the difficulty of segmentation,especially in the boundary area of the targets.Experiments show that the new model in this paper performs better than the classic U-Net in this respect.