Research And System Implementation Of Cell Image Classification Based On Swin-Cell Transformer

In the clinical diagnosis of blood diseases,the analysis of bone marrow cells is a very important part of the diagnosis and treatment process.Moreover,the morphology of the diseased white blood cells is often more complex,so the morphological judgment of the diseased cells becomes the priority among priorities.At present,the detection of bone marrow cells in the hospital mainly includes the use of blood cell analyzer analysis and manual microscopy.However,manual microscopy requires a large amount of labor cost,and has a great test on the visual resolution and the experience of doctors.However,the accuracy of blood cell analyzer on the market is quite different from that of manual microscopy,and some blood samples need to be further manual reexamination.In addition to the difficulties of blood cell analyzer and manual microscopy in the examination of bone marrow cells,the performance of traditional convolutional neural networks in the classification of bone marrow cells has also reached a bottleneck.Although convolutional neural network has translation invariance,scale invariance,rotation invariance and other excellent characteristics in the field of computer vision,with the continuous integration of natural language processing technology and computer vision technology,under the task of classifying large data sets,The model of Transformer class transfers the model advantages in natural language processing to the traditional computer vision model,and the model performance after integration has reached a new height.Therefore,it has become a new trend to use Transformer class model to deal with image classification.The main contributions of this thesis are as follows:Created a data set of PBC-6 diseased cells to support the diagnosis of diseases such as leukemia.The data set included a total of about 11,000 cell images of six kinds of diseased cells including toxic granules,vacuoles,and Durer’s bodies.The samples in the data set were processed by filtering methods such as median filtering and Gaussian pyramid,as well as enhancement methods such as rotation and flip.The classification performance of the models was further improved by the models and optimization of the training methods.In the classification task of conventional bone marrow cells,the overall test accuracy of Swin-Cell Transformer was 0.90,and the test accuracy of the Res Net50 model used by the original authors of the dataset was 0.82.The model proposed in this paper improved the test accuracy in the task of conventional bone marrow cell classification by 8 percentage points.The classification of lesioned leukocytes is a relatively new application scenario,and this paper also used the four classification models mentioned above to conduct classification experiments on a private dataset,and the accuracy of the Swin-Cell Transformer model for classification was 0.86.Based on the Swin-Cell Transformer,we developed the software for assisting the diagnosis and treatment of bone marrow cells.In this paper,a complementary diagnosis and treatment system for myeloid cell diseases was developed based on the Swin-Cell Transformer prediction model.The system consists of five modules: electronic medical record management,prediction of leukemia and other related diseases,examination system,data statistics,and authority management.The dynamic parameters and infinite parameter space characteristics of the Swin-Cell Transformer model proposed in this paper as well as the custom window mechanism during training effectively enhance the information interaction in the feature graph,and its performance is better than that of the traditional convolutional neural network and other Transformer models.It has greatly improved the effect of medical cell image classification task.