Construction Of Transfer Learning Models Based On Plasma Cell Morphology To Identify Complex Karyotypes In Multiple Myeloma

Background:Multiple myeloma（MM）is a monoclonal proliferative hematologic malignancy of plasma cells.The median age of onset is about 65-70 years old.The incidence of males is higher than that of females,about 1.4:1.The main mechanism of MM is that plasma cells in the bone marrow secrete a large number of monoclonal immunoglobulins after malignant proliferation,thereby causing damage to multiple organs and tissues of the body.Its clinical manifestations are often anemia,renal function damage,bone destruction and so on.The overall survival period of MM is about 2-10 years.MM prognostic evaluation standards mainly include Internatioal staging system（ISS）standard,Revised-international staging system（R-ISS）standard,International Myeloma Working Group（IMWG）risk staging system,and m SMART3.0 staging system.To be improved.The cytological diagnosis of MM mainly relies on the myeloma cells in the bone marrow smear under the counting microscope.For experts in pathology and bone marrow cell morphology,this process is tedious,time-consuming,and difficult to standardize.In addition,the artificial visual diagnosis is subject to a certain degree,which may increase the risk of error in the final decision.Due to the high genetic instability of MM,abnormal karyotypes are found in about 30%-50% of MM cases.Studies have shown that complex karyotype changes often indicate poor prognosis for MM patients.Therefore,karyotype detection has a certain value in assisting prognosis and guiding treatment strategies.However,due to the low proliferative activity of malignant plasma cells in vitro,the study of cytogenetic abnormalities by karyotyping has certain limitations.Especially for complex karyotype changes involving multiple chromosomal complex variations,fluorescence in situ hybridization（FISH）,which is widely used in clinical practice,cannot completely detect all karyotype changes in MM chromosomes.There are still limitations in terms of cytogenetic alterations.Studies have shown that the morphological characteristics of MM tumor plasma cells have certain clinical prognostic value,but there is no research on the correlation between complex karyotype and cell morphology.Based on the above background,the existing MM prognosis judgment models still face many challenges,and the clinical prognosis judgment of MM is not perfect.In recent years,artificial intelligence deep learning has made breakthroughs in automatic feature learning by imitating the structure and operation of the human brain.In particular,advances in computer-aided methods have enabled faster and more repeatable medical image analysis than manual analysis.Therefore,it has attracted great attention in computer vision tasks.Therefore,we propose a new method to use the cell morphology to judge the complex karyotype of MM deep learning model to assist the clinical diagnosis and treatment process.Objective:The computer deep learning model was used to identify the morphology of bone marrow plasma cells with complex karyotypes in MM.And utilizing the characteristics of complex karyotype MM patients’ bone marrow plasma cells,deep learning modle has capacity of co-diagnosis in judging the clinical prognosis and assist in guiding the treatment strategy in future.Materials and Methods:The bone marrow smears are from MM patients who were treated in China-Japan Union Hospital Affiliated to Jilin University during the period from September 2016 to December 2021.In this study,karyotype analysis defined unrelated karyotype changes with 3 or more abnormalities as complex karyotype MM,and unrelated karyotype changes with less than 3 abnormalities as uncomplicated karyotype MM.According to the karyotype analysis report of MM patients,the myeloma plasma cells were divided into two groups.There were 39 cases with uncomplicated karyotype and 40 cases with uncomplicated karyotype.Two bone marrow smear specimens were selected from each patient,and the bone marrow smear specimens were examined by experienced bone marrow cytomorphology experts with an OLYMPUS BX51 microscope at 10×100magnification.And we use MATLAB algorithms to segment cells.Finally we get a dataset 12,000 MM bone marrow plasma cells,which contains 6000 complex karyotypes cells and 6000 non-complex karyotype plasma cells.Applying MATLAB for transfer deep learning,ImageNet-based pre-trained architectures include AlexNet,ResNet 50,Inception ResNet v2,and VGG16.Results:In this study,we have constructed four artificial intelligence（AI）transfer learning models algorithms.AlexNet model’s accuracy is 90.90% and the area under the ROC curve（AUC）is 0.9679.ResNet 50 model’s accuracy is 94.31% and AUC is 0.9857.The Inception ResNet v2 model’s accuracy is 87.51% and AUC is 0.9566.The VGG16model’s accuracy is 93.87% and AUC is 0.9860.Conclusions:1.Our study provides a large database of original images of MM bone marrow plasma cells,which are from MM patients’ bone marrow smears.And our study will provide certain data support for future follow-up studies based on plasma cells.2.Our study proposes four AI transfer learning algorithms to realize the identification of complex karyotype MM bone marrow plasma cells.The artificial intelligence transfer learning algorithms can quickly and accurately predict bone marrow plasma cells of complex karyotype MM patients.This study has value for clinical prognosis judgment and clinical individualized precision treatment.3.In this study,we have constructed four deep learning classifier models,which are AlexNet,ResNet 50,Inception ResNet v2,and VGG16.Each of the deep learning classifier models has achieved good classification and prediction results.And ResNet50 has the highest accuracy in the four models.