Research And Implementation Of Key Algorithms For Microscopic Image Analysis Based On Deep Learning

With the rapid development of computer technology,deep learning has made great breakthroughs in the field of medical image recognition.The medical case studied in this paper is diffuse axonal injury.The picture is a microscopic image of brainstem slice.of mouses or hippocampals The research goal is to use deep learning to detect the number and position of restraction balls in the slice image to determine the disease.The key algorithm used in this research is Faster RCNN object detection algorithm.Faster RCNN is mainly composed of feature extraction network,region propasal network,ROI pooling network,and fully connected network.Faster RCNN has the advantages of fast training speed,high detection accuracys and simple training step.The data set of this experiment is microscopic slice images,which has the characteristics of small detection objects,few samples,and high resolution.However the original data set is the pascal＿voc data set,which has the characteristics of big detection objects,lots of samples and low resolution.Based on the differences in the data sets,this paper improves the network structure of Faster RCNN and personalizes the hyperparameters to improve the accuracy and speed of the model detection on the microscopic image data set.The disadvantages of the original network structure are as follows: Use VGG16 as the feature extraction layer,the ability of which to extract microslice image features is insufficient,and the training weight is too large,which will cause slow training speed;the region propasal network(RPN)will randomly generate the region of interests(Rois),which will to a certain extent result in low training efficiency.The improvement of the network structure are as follows: A deeper residual network(Res Net)with stronger ability to extract features is used to replace the traditional VGG16 convolutional neural network.Res Net has better feature extraction for images,and has fewer training weights,which is easy to train;Adopting online hard example mining algorithms to improve the way that RPN networks generate Rois can allow the model to be more fully trained,thereby improving detection accuracy.The original hyperparameters are set for the pascal voc dataset but not suitable for microscopic image datasets.The experiment personalizes the hyperparameters from four aspects including the training process,test process,loss function,and anchor generation to improve detection accuracy and speed.The accuracy of the original algorithm on the microscopic image dataset is 90.62%,however the accuracy of the improved Faster RCNN is 94.73%.so the accuracy is improved by about 4%,and the test speed is improved about 0.1s.