Medical Image Segmentation Based On Deep Learning

With the general applications of medical images in the field of medical care,the segmentation of medical images plays an important role in pathological analysis,clinical diagnosis,and medical intervention.In recent years,with the development of feature extraction ability of convolutional neural networks,more and more deep learning technologies have been applied to medical image segmentation,and have achieved better results compared with the traditional segmentation methods.Full convolutional neural network(FCN)is a deep learning technique widely used in image segmentation.However,due to the complexity of the size and shape of the segmentation objects and the unbalanced distribution of positive and negative samples in medical images,FCN has not achieved a good expected performance in some medical image segmentation tasks.In order to achieve more accurate segmentation results,this thesis proposes two improved algorithms based on the full convolutional neural networks.The proposed algorithms have good robustness to the morphological changes of the segmentation objects.It can solve the problem caused by unbalanced distribution of positive and negative samples,and improve the segmentation accuracy.The main work and conclusions of this thesis are as follows:1.Inspired by DenseNet and U-Net,this thesis proposes a novel medical image segmentation method based on fully convolutional DenseNet.Based on the popular deep learning theory,the method loads DenseNet from natural images to medical image datasets.The full convolution DenseNet uses deconvolutions and the connection way similar to U-Net,and achieves end to end image segmentation.Furthermore,improving the Dice similarity loss function can solve the problems that the proportion of background in medical images is much larger than that of object region and some pixels are difficult to be accurately identify.The experiment on prostate segmentation dataset shows that the proposed method is more effective and takes less time compared with other main methods.2.This thesis develops an improved U-Net with deformable encoder and reshaping upsampling convolution decoder.The method applies and improves deformable convolutions to enhance the learning ability of the encoder for geometric transformations.A novel upsampling method named reshape upsampling convolution(RUC)is proposed for better restoring resolution and fusion features.Futhermore,focal loss function is used to medical image segmentation tasks for solving the problem of model overwhelmed caused by class-imbalance and too many simple samples in biomedical images.The model not only reduces the number of parameters of U-Net,but also achieves good results on Drosophila EM dataset and Warwick-QU dataset.