Differentiation Of Multiple Myeloma And Systemic Light Chain Amyloidosis By Artificial Intelligence Blood Cell Morphological Analysis

PurposeIn recent years,artificial intelligence has been widely used in the field of blood cell morphology analysis,which has promoted the development of the diagnosis,identification and prognosis of diseases of the blood tumor system.The identification of different plasma cell clonal diseases mainly depends on clinical manifestations,laboratory,imaging and other examinations.At present,the number of clonal plasma cells in the bone marrow cannot be used to identify systemic light chain amyloidosis（AL）and multiple myeloma（MM）,as some AL patients also show high tumor burden in the bone marrow.The difference between plasma cells itself may also be one of the reasons for the occurrence of different types of plasma cell diseases.Therefore,based on artificial intelligence,we established multiple myeloma and systemic light chain amyloid based on the characteristics of plasma cell morphology and nuclear morphology in bone marrow smears.Degeneration differentially diagnoses models and evaluates model performance.The aim of this study was to differentiate between the two diseases by the morphological features of plasma cells and nuclei.For patients diagnosed with plasma cell malignancies by bone marrow examination,it is expected to identify and classify the plasma cell morphology and plasma cell nuclear morphology characteristics in the bone marrow through artificial intelligence,so as to improve the efficiency of disease diagnosis.Materials and MethodsWe collected and collated the bone marrow smears of 10 patients with systemic light chain amyloidosis and 10 patients with multiple myeloma who had been diagnosed in the China-Japan Union Hospital of Jilin University from 2011 to 2021.Each patient choose two bone marrow smears.Using microscope and MATLAB image acquisition toolbox for cell image acquisition,a database of bone marrow cell morphology in systemic light chain amyloidosis（AL）and multiple myeloma（MM）was established.Cells were labeled by two bone marrow pathologists,and 7500 images of AL plasma cells and 8000 images of MM plasma cells were obtained by Image J segmentation,and a single plasma cell morphology data set was established.The features of nuclei were extracted by Image J segmentation,and R language was used to build a classical machine learning model based on the morphological features of plasma cell nuclei for AL and MM discrimination.The model performance was evaluated by ROC curve（Receiver operating characteristic curve,ROC curve for short）and AUC（Area Under the Curve,AUC）.The MATLAB deep learning toolbox pre-trained network model was used to train plasma cell morphology.The pre-trained network models included Vgg16,Vgg19,Googlenet,Alexnet,and Resnet18.Evaluate model performance by confusion matrix of accuracy,precision,AUC,recall,etc.Results1.In this study,a morphological database of AL and MM bone marrow cells was established,7500 AL plasma cell images and 8000 MM plasma cell images were obtained by Image J segmentation,a single plasma cell image data set was established,and the morphological characteristics of plasma cell nuclei were extracted and obtained the Summary Table.2.In this study,a classic machine learning model was constructed to identify AL and MM based on the morphological characteristics of plasma cell nucleus.According to the summary data,the AUC value of the area under the ROC curve of the model was obtained.The AUC value of the linear regression model is 68%,and the AUC of the SVM model is 79%,the AUC value of the random forest model is84%,and the AUC value of the boost model is 85%.The Boost model confusion matrix has an accuracy of 79.7%,a total error rate of 20.3%,and an average classification error rate of 27.35%.3.In this study,a deep learning model was constructed to discriminate AL and MM based on plasma cell morphology.The accuracy of the Alexnet model is89.97%,and the area under the ROC curve AUC value is 87.51%.The accuracy of the Resnet18 model is 90.59%,and the AUC value is 98.18%.The accuracy of Googlenet is 77.15% and the AUC value is 94.27%.The accuracy of the Vgg16 model is 89.83%,and the AUC value is 88.22%.The accuracy of the Vgg19 model is 89.15%,and the AUC value is 85.49%.Conclusions1.This study uses a dataset of 15,500 plasma cell images to evaluate the classification performance.The deep learning algorithms used in the evaluation work well,providing image data for future research based on plasma cell and nuclear morphology.2.The classical machine learning model constructed based on plasma cell nuclear features in this study can better classify and identify AL and MM.Compared with linear regression,SVM,and random forest,the Boost model based on plasma cell nuclear morphological features has the best performance in identifying AL and MM.However,the average error rate of the confusion matrix is high,which may have certain difficulties in clinical application.3.In this study,the five deep learning models Vgg16,Vgg19,Googlenet,Alexnet,and Resnet18 are constructed based on the morphology of plasma cells can better classify and identify AL and MM,and through comparison,it is found that the Resnet18 model has the best performance and can pass the plasma cell well.The cell morphology can realize the identification of the two diseases,and provides meaningful clues for the identification of plasma cell malignancies.