Session: 653. Multiple Myeloma: Clinical and Epidemiological: Poster II
Hematology Disease Topics & Pathways:
Research, Artificial intelligence (AI), Diseases, Technology and Procedures
Bone marrow smears obtained from patients diagnosed with MM and AL in hospitals were retrospectively analyzed. The bone marrow cell images under Olympus BX51 with oil lens resolution (1000X) were obtained via the Windows camera and the Imagetion Toolbox acquisition kit of MATLAB. Single plasma cells were extracted via ImageJ/MATLAB to construct the dataset. Using the plasma cell dataset, the deep learning models used to identify MM and AL was trained on the ImageNet pretraining network models shufflenet, ResNet50, ResNet101, SqueezeNet, Inceptionv3, and MobileNetv2. The ratio of the training set ,the validation set and the test set is set to 7:2:1. The confusion matrix, receiver operating characteristic (ROC) curve and the area under the ROC curve (AUC) were used to evaluate the deep learning models.
We built a single plasma cell dataset of 10000+ which contains data from MM and AL. And we generated deep learning models on the basis of the pretrained network models of shufflenet, ResNet50, ResNet101, SqueezeNet, Inceptionv3, and MobileNetv2. By testing deep learning models, we obtained the confusion matrix of each model and calculated the accuracy, AUC, precision and recall. We found that the confusion matrix shows that most deep learning models misjudge only a small number of AL plasma cells as MM plasma cells. The accuracy of the deep learning models in identifying AL and MM plasma cells are above 0.97. The accuracy rates of shufflenet, ResNet50, ResNet101, SqueezeNet, Inceptionv3 and MobileNetv2 were 0.99, 0.99, 1.00, 0.99, 0.97 and 0.99, respectively. The AUC values were all above 0.95, and the AUC values of the six pretrained models were 0.9949, 0.9936, 0.9962, 0.9949, 0.9720 and 0.99. The accuracies of the six pretraining models are 0.99, 0.99, 0.99, 0.99, 0.95, and 0.9907. The recall rates of the six pretrained models are 1.00, 1.00, 1.00, 1.00, and 0.99. Among them, the accuracy, accuracy and recall rate of ResNet101 are the highest, are 1.00, 0.99 and 1.00, respectively. It has the best performance. Therefore, we concluded that The deep learning models based on plasma cell morphology could identify MM and AL plasma cells accurately.
Disclosures: No relevant conflicts of interest to declare.
See more of: Oral and Poster Abstracts