Research Of Deep Learning In Recognition Of Placental Ultrasound Image And Dermoscopy Image

Currently,clinicians rely on the observation and analysis of medical images for the diagnosis of many diseases.However,due to the limitations of the imaging quality of medical instruments and the subjective errors of clinicians during analysis,misdiagnosis or missed diagnosis may occur.In order to solve this problem,image automatic analysis methods based on computer-aided diagnosis have been attracting attention.In recent years,deep learning has been continuously introduced into various applications in the medical imaging field,such as disease classification,prediction,lesion segmentation,etc.,due to its powerful self-learning ability,which has achieved great success.Compared with traditional machine learning methods,deep learning methods can eliminate the complexity of feature engineering,and efficiently learn the target features through convolution operations to complete specific tasks.However,its application in medical imaging still faces many difficulties.The amount of medical images is too small to train the model well which may result in poor performance of the model;Imaging instruments differ greatly in imaging results due to different hardware or parameters,and some instruments have poor imaging quality.These factors further increase the difficulty of model optimization,and require high generalization ability of the model.Focusing on above problems,this paper explores the application of deep learning methods in the medical image field based on placental ultrasound images and dermoscopic images.The main research results include automatic placental maturity grading based on hybrid descriptors,convolutional descriptors aggregation via multi-network for melanoma recognition and automated skin lesion segmentation via generative adversarial networks with dual discrimination strategy.In the study of placental maturity grading,this paper mainly proposes an ensemble model,which combines convolutional features in multiple convolutional neural networks and manual features,and then performs Fisher vector encoding to achieve automatic grading.This model both takes into account the rich local information in manual features and the deep information contained in the convolutional features,which complements each other to better express the image.To solve the problem of insufficient data,the paper expands the training set via data augmentation and adopts the transfer learning strategy to prevent overfitting.A large number of experimental results confirm the effectiveness of the proposed method.For the recognition of dermoscopy melanoma,a multi-network framework based on hybrid coding is proposed.The framework extracts the deep features of several different types of convolutional neural networks that have complemental information.Then,the convolutional descriptor aggregation method is used to further select and fuse these features of each network,so that the feature expression with rich information and less redundancy is obtained,which can reduce computational burden while ensuring the feature is effective.The fused features are then encoded using the Fisher vector.This model was validated on the ISBI 2016 and ISBI 2017 datasets and achieved good results.In addition,this paper also presents a skin lesion segmentation method based on generative adversarial networks with dual discrimination strategy.By adding a dense atrous convolution block to the generator to get more fine-grained information and expand the receptive field.Two discriminators are used in the discrimination module to constrain the output of the generator to achieve accurate segmentation of skin lesion.Compared with other segmentation networks,proposed network can learn more image details and has stronger supervision.The model was validated in the ISIC 2017 dataset and achieved better results.