Image Classification Of Acute Lymphoblastic Leukemia Based On Deep Learning

Acute lymphoblastic leukemia is a kind of acute cancer,and the main ill population is children aged 0-9.If they are not found and treated in time,they will die within a few months.Therefore,the early diagnosis of leukemia is very important for the treatment of patients.However,due to the highly similar morphology between normal cells and cancer cells,it is particularly difficult for pathologists to diagnose.Although there are instruments,such as flow cytometry that can assist doctors in diagnosis,the expensive price makes it difficult to be widely applied.With the improvement of hardware and the development and application of deep learning technology in recent years,computer-aided diagnosis of doctors has become possible.Its advantages are that it can greatly reduce the workload of pathologists,and its low cost,so it can be widely used in underdeveloped medical areas.Based on the investigation of the research status of deep learning technology,we proposes a method of WBC(White Blood Cell)image classification based on deep learning.The purpose of this method is to build convolutional neural network by using deep learning knowledge to complete the automatic classification of leukocyte images.The main work and innovation of this paper include:(1)We improved the training method by using the strategies of transfer learning,data augmentation and cosine decay to speed up the convergence of the network,expand the size of the data set and avoid the model falling into the local minimum.In addition,different learning rates are set for different layers of the network to make the best of transfer learning.According to these methods,the accuracy and F1 score of the model are improved by 9.8%and 5.4%respectively.(2)A preprocessing method of mixed multiscale image is proposed.Linear interpolation is used to mix images of different scales,so that the feature map can contain local and global features at the same time,which alleviates the fine-grained problem.The experimental results show that the accuracy and F1 score of the model are improved by 2.3%and 1.0%,respectively.(3)On the basis of ResNet network,we propose a multi branch pool layer of resnet-50 network and add attention mechanism,termed as MBPA-ResNet,which makes more effective use of the mixed multi-scale image method proposed in this paper.In this paper,a large number of comparative experiments are designed for the proposed method.The experimental results show that the accuracy of the model is improved from the accuracy of 0.742 and the F1-score of 0.828 to 0.878 and 0.901,which verifies the superiority of this method.