Application of Transfer Learning to Recognize Early Death Myeloma Patients Based on Plasma Cell Morphology

Multiple myeloma (MM) is a hematological malignancy that caused 187,952 new cases and 121,388 deaths worldwide in 2020. Over the past 10 to 15 years, with the advent of new drugs and technologies, the overall survival of MM patients has been extended from about three years to more than five years. However, not all patients with MM benefit from treatment. Some patients still die within 12 months of diagnosis, known as early mortality (EM), which severely impedes further survival in patients with MM. There are still great challenges in how to identify the high-risk group of EM at the initial diagnosis. Previous studies have confirmed that the degree of plasma cell maturity in bone marrow smears is related to the prognosis of MM, but there are few studies on whether the prognosis can be determined by plasma cell morphology in bone marrow smears. Morphological examination of bone marrow cells is the cornerstone of the diagnosis and treatment of various hematological diseases. Although there are many approaches, including cytogenetics, immunophenotyping, and molecular genetics, morphological examination remains the basic item and the first critical step in diagnosis and therapeutic response monitoring. However, at present, bone marrow smears largely rely on human recognition, which is time-consuming, and the accuracy is easily affected by subjective factors of the examiner. So human recognition cannot be used in large-scale judgment of MM prognosis. Therefore, we hypothesize that we can use the recognition of plasma cell morphology in bone marrow smears to determine early death, and we can introduce artificial intelligence into this process. We attempted to establish deep learning models to distinguish between groups with early death and good prognosis based on bone marrow plasma cell images and evaluate its effectiveness.

We retrospectively analyzed the bone marrow smears of patients admitted to affiliated hospitals with a definite diagnosis of MM, and divided them into EM group and good prognosis group. Full field plasma cell images (resolution 1920x1080) containing plasma cells were acquired with Windows camera and Marketing Toolbox in MATLAB to establish a marrow cell dataset. MATLAB's Imagelabeler Toolbox segment and label plasma cells from MM bone marrow smears to obtain a single cell level plasma cell dataset with 4677 early dead plasma cells and 7977 non-early dead plasma cells. The patient's bone marrow smear specimens were divided into three groups according to 7:2:1 for training, validation and testing the deep learning models. Deep learning models were trained based on ImageNet pre-trained network models shufflenet, resnet50, resnet101, SqueezeNet, inceptionv3, and mobilenetv2 to distinguish EM from good prognosis. Deep learning models were evaluated through the confusion matrix, ROC(receiver operating characteristic) curve and AUC(area under the curve, the area under the ROC curve) value.

We finally established a 5000+MM single plasma cell dataset and generated six deep learning models based on the pre-trained network models to judge patient prognosis. By testing six deep learning models, the confusion matrix of the six models were obtained. By evaluating the test results, we found that shufflenet, mobilenetv2 and resnet101 performed well, with relatively few errors in classification and accuracy rates of 0.80, 0.78 and 0.87, respectively, all above 0.75. The area under ROC curve (AUC values) were close to or above 0.9, which were 0.89, 0.87 and 0.94, respectively. The performance of resnet50, squeezenet, and inceptionv3 was not satisfactory, and the area under the ROC curve (AUC values) did not exceed 0.5, which were 0.49, 0.5, and 0.47, respectively. Further adjustment and further training were needed. The appearance of this result may be related to the structure of the models and the parameters, and the ability of these models to recognize EM may be improved by the later adjustment of parameters. At the same time, shufflenet, mobilenetv2 and resnet101 models can achieve high accuracy through training, which also shows that it is feasible to use artificial intelligence to identify the morphology of plasma cells in the bone marrow smear and further predict the prognosis of patients. In the future, with the introduction of new models, the technology will be more mature, so that medical staff can predict EM and provide personalized treatment.

No relevant conflicts of interest to declare.