Recognition Of Human Protein Atlas Image Based On SSIP And AMC-Net

Due to the complex layout of multi-label images,occlusion and poor performance of single-label image algorithm in direct migration,it is difficult for the multi-label image recognition model to achieve the requirements of practical application.In addition,medical images are difficult to process due to such factors as small feature differences among different categories,high feature recognition difficulty and unbalanced data.Deep convolutional neural network(DNN)has been widely used in single label image recognition because of its strong fitting ability and good generalization ability.However,the difficulty of the multi-label image task is much higher than that of the single-label image task,it is invalid to transfer the excellent models and methods in the single-label recognition task to the multi-label recognition task.Aiming at the human protein atlas image dataset in the field of medical image processing,this thesis analyzes the relevant theories such as target detection,image segmentation and the particularity of the dataset.Based on the selective search algorithm,we propose the combined selective search algorithm.Meanwhile,based on multi-label problem transformation and multi-scale model,the depth model SSIP and the convolutional network AMC-Net are proposed respectively.In addition,some optimization methods are proposed.The research results of this thesis are summarized as follows:(1)AMC-Net: a multi-input asymmetric and multi-scale convolutional neural network,which called AMC-Net,is proposed to solve the problem of large variation of protein feature scales.Images of the same protein that are scaled to three different scales are inputs to AMC-Net.The three branches of the model have different convolution kernel size and pooling kernel size,and the large-scale image corresponds to the large-scale convolution kernel branch.Through experiments,we have proved that this multi-scale model can achieve a good classification effect and obtain an F1 score of 0.821.(2)SSIP: aiming at the multi-label attribute of dataset,we design a novel end-to-end multi-label classification framework,called SSIP,which integrates our improved detection algorithm,convolutional neural network,max-pooling layer and threshold method.The detection algorithm in SSIP,combined selective search algorithm,is based on selective search algorithm,and is proposed in view of the particularity of the dataset information stored in different channels.In addition,we also conducted a large number of experiments to evaluate the SSIP framework and proved its robustness,in which the F1-score reached 0.83.(3)Model optimization: in terms of monitoring signal optimization,this thesis introduces the combined loss function,which is composed of the binary cross-entropy loss function and F1-score loss function.While alleviating the problem of overfitting caused by the serious imbalance of datasets,our function can guide the model to improve F1-score without extra computation(about 0.015 improvement over the focal loss).On the optimization of object detection in SSIP framework,this thesis proposes a multi-scale object detection method,which effectively reduces the missed detection rate and improves the F1-score by about 0.016.In this thesis,the human protein atlas dataset was taken as the experimental object.Through a large number of comparative experiments,the robustness of the proposed AMC-net and SSIP were proved.Meanwhile,it is proved that the proposed loss function optimization and multi-scale optimization also significantly improve the performance of the deep model.