Study Of Deep Learning And Applications In Medical Image Analysis

With the improved quality of medical images,the rise of medical big data,more accu-rate and robust algorithms,especially the extensive use of deep learning algorithms in the field of medical AI,the previous difficult computer-aided medical image process-ing has made unprecedented progress.Medical image processing algorithms based on deep learning make computer-aided medical diagnosis possible in clinical practice.In this paper,we focus on two very challenging applications in medical image analysis:"Lung cancer pathological cell image recognition" and"prostate CT image segmen-tation",analyzing the current existing problems,and trying to solve the problem with deep learning techniques,especially based on convolutional neural networks.We pro-pose the following algorithms:(1)A multi-instance deep convolutional network based on prototype learning is proposed.In lung cancer images,cancer cells tend to occupy only a small area.When using a general classifier for cancer diagnosis,they often suffer from a lot of back-ground noise.The traditional convolutional neural network is a fully supervised deep learning model that requires the complete sample labels.The application of convo-lutional neural networks is greatly constrained in the practical application of weak supervision such as missing cell labels.To solve the multi-instance learning problem in weak supervised condition,a new multi-instance deep convolutional network model is proposed.The model introduces a new prototype learning layer.This layer uses a prototype-based similarity-based algorithm to implement mapping of instance features to bag features,enabling the network to give class labels at the bag level,thus com-pleting the learning process for the entire model.Thereby solving the problem of deep learning method under weak supervision.(2)A fully convolutional network based on distinctive curve guidance is proposed.In order to do radiotherapy of prostate cancer,it is a very challenging task to accurate-ly delineate pelvic organs(including prostate,bladder and rectum)in CT images by computer algorithms.We propose a two-stage deep learning-based approach that in-cludes a novel distinctive curve guided fully convolutional network(FCN),to address the problems of low soft tissue contrast in CT images,and large shapes and appearance variations of pelvic organs.We introduced a two-stage framework that firstly locate the relatively small target organ region in the raw CT image,allowing the network to better segment the organ.To help identify the boundaries of these undistinguished pelvic organ boundaries,we use a distinctive curve to guide a multi-task FCN for fine segmentation.In the experiment we verified that the method is accurate and robust in the segmentation task of pelvic organs and outperformed the state-of-the-art methods.(3)A hierarchically fused multi-task fully convolution network,namely HFFCN,is proposed.Conventional multi-task deep networks often share most of their layers and parameters between different tasks,often limiting their ability to suit for data.In order to solve this problem,we propose a hierarchically fused multi-task full convolu-tional neural network,namely HFFCN.It goes as a Y-shaped structure that allows the network to share an encoding path,but learns task-specific information in two sepa-rate decoding path.Then,we proposed a new type of Information Sharing(IS)block to provide communication and information sharing between the two decoding path-s.This allows our HFFCN network to better learn the hierarchical representation of information between tasks,and at the same time to preserve the task independent rep-resentations.Experiments show that our approach can outperform the conventional multi-tasking deep network architectures and the state-of-the-art methods.