Application Of Deep Convolutional Neural Network In Medical Image Segmentation

Image segmentation is an important research topic in medical image processing and analysis.However,currently,it highly relies on the doctor’s manual segmentation,which is not only time-consuming and labor-intensive but also suffering from high subjective errors.To solve these issues,medical image processing methods based on computer-aided diagnosis have been attracting attention.Deep learning methods have been achieving great success in the computer vision field,and thus are introduced to medical image analysis field too.With deep learning methods’ powerful feature learning capabilities,they have been achieving success in disease diagnosis,prediction,tumor detection,and segmentation.However,there are still many challenges in applying deep learning methods to medical image segmentation.For example,1.Deep learning methods acquire big data to train a state-of-the-art model while medical image datasets are usually small,which leads to poor generalization;2.As for the ultrasound image segmentation task,it’s a great challenge to train a robust model due to ultrasound images’ poor image quality with noise and artifacts;3.As for histopathology image segmentation task,images are usually high resolution,and it’s tough for deep learning model to train and learn discriminative feature due to limited computing resources.Focusing on the above challenges,this thesis explores the applications of deep learning methods for medical image segmentation based on echocardiography images,prostate cancer histopathology images,and CT images from head & neck and chest.Our main research results include an attention-guided dual-path network for pediatric echocardiography segmentation,Gleason grading of prostate cancer from histopathology images using PSPNet,and automatic structure segmentation for radiotherapy planning using nnU-Net.1.In the study of pediatric echocardiography segmentation,a novel attention-guided dualpath network is proposed.We first extract features from different depths with a dual-path network structure,and an attention mechanism module is added to refine learned features,improves the representative capability of the neural network,and finally boost the segmentation performance of echocardiography images.Many experiments using a self-collected pediatric echocardiography dataset and a public adult echocardiography dataset are designed and conduct,and the results confirm the performance of the proposed method.2.In the study of Gleason grading of prostate cancer,a segmentation network base on PSPNet is used for this task.We build PSPNet with ResNet 101 as the backbone to extract features and then refine features from different scale via pyramid pooling module,we also add an auxiliary branch to boost the segmentation performance.We obtain pixel-level Gleason grade of histopathology images with segmentation,and then calculate the core-level Gleason grade according to the Gleason grade system.Experiments on the public MICCAI Gleason 2019 dataset confirm the robustness and effectiveness of the proposed method.3.In the study of radiotherapy planning,nnU-Net is used to segment structures from 3D CT images in radiotherapy planning.Building from the na?ve U-Net,nnU-Net automatically adapts itself according to the property of the given dataset.These adaptions include the preprocessing scheme,U-Net network structure,and training strategy.Unlike those networks with complicated structures,nnU-Net has the power to design a simple but suitable U-Net for the given dataset.Experiments on the public MICCAI StructSeg 2019 dataset confirm the robustness and effectiveness of the proposed method.