Research On Cell Classification Method Based On Peripheral Blood Smears Of Patients With Acute Myeloid Leukemia

Leukaemia is a highly lethal blood disorder and the main detection methods are manual microscopy and blood cell analysers,but both are highly repetitive,laborious,time-consuming or expensive.At present,blood cell detection by microscopic images is still an important means of diagnosing leukaemia.Mainstream image processing techniques bring great advantages to the classification of blood cell images,which can not only process the morphological information of cells,but also be efficient.In this regard,this thesis investigates the classification and identification of cells in peripheral blood smears of acute myeloid leukaemia based on the images of these cells.Four deep learning image classification models,Alex Net,Google Net,VGG16 and Res Net50,were used to classify 15 types of cells,and the classification accuracy was compared and analysed with the evaluation index.In order to verify the effect of different depths of Res Net models on the experimental results,experiments were conducted on three different depths of Res Net34,Res Net50 and Res Net101 models.The results show that the classification accuracy of Res Net50 is 0.919,which has a clear advantage over Res Net34’s0.861 and Res Net101’s 0.850.In order to further improve the classification results,and to address the situation that the morphology of cell images is similar between classes but varies greatly within classes,two methods of fine-grained classification are proposed from the perspective of fine-grained classification.One approach is to highlight non-significant but beneficial features by suppressing significant features and thus highlighting non-significant but beneficial features for classification.A feature enhancement suppression module,a feature diversification module and an attention mechanism module are introduced to further improve the generalisation ability of the model by enhancing non-significant information that is beneficial for classification.The experimental results show that after adding the feature suppression and enhancement module and the feature diversification module,the classification accuracy of the model is 0.927,an improvement of 0.8% compared with the original Res Net50 model,indicating that the model can perform better classification.However,because the first method is not superior in classifying four classes of cells with large differences in appearance and morphology,a second structure is proposed,which is to obtain multi-scale patch groups by adding a local stream,first inputting the images into the pyramidal convolution,randomly selecting the corresponding patches from each scale patch group,then iterating the patch feature information through the LSTM network,and then using the attention mechanism to The global stream of the main branch can extract the global features of the image,while the local stream extracts the finer features,and the two groups of features are combined in a complementary way.The experimental results show that after adding the local stream branch,the classification accuracy of the model is 0.939,which is a 2% improvement compared with the original base network,indicating that the model can further improve the classification accuracy.