Few-Shot Learning using Siamese Twin Network for the Classification of Blood Cells

Abstract

Automated classification of blood cells from microscopic images is an interesting research area owing to advancements of efficient neural network architectures. Here, we developed a few-shot contrastive learning model for the classification of peripheral blood cells including lymphocytes, monocytes, basophils, eosinophils, neutrophils, immature granulocytes, erythroblasts, and platelets using EfficientNet as a base model and contrastive loss as a loss function. A total of 17092 publicly accessible images acquired using the CellaVision DM96 were analyzed. From 125 images of each cell type, 20000 image pairs are generated for Siamese twin network (STN) training and another 125 images from each cell type are used for few-shot validation. Therefore, out of 17092 images, 6% were used for training, 6% for few-shot validation and rest 88% for few-shot testing. This architecture demonstrates an overall accuracy of 97.21% during 8-way 3-shot testing for the classification of all cell types with an accuracy of 97.72% for the classification of white blood cells alone. Further, we propose a novel class activation mapping scheme for the interpretability of the model decisions suitable for STN. To conclude, the proposed framework based on contrastive learning could be used for the fully automated self-exploratory classification and identification of peripheral blood cells.