Localization and recognition of leukocytes in peripheral blood: A deep learning approach

Automatic recognition and classification of leukocytes helps medical practitioners to diagnose various blood-related diseases by analysing their percentages. Different researchers have come up with different algorithms that use traditional learning for the classification of different types of leukocytes. In contrast to traditional learning, in which no knowledge is retained that can be transferred from one model to another, our proposed algorithm uses deep learning approach for segmentation and classification. The proposed algorithm has two-stage pipelining consisting of semantic segmentation and transfer learning-based classification. Here, we have used pre-trained networks, utilizing knowledge from previously learned tasks, called DeepLabv3+ for segmentation of leukocytes and AlexNet to classify five categories of leukocytes in peripheral blood from whole blood smear microscopic images. For experimentation, a microscopic blood image dataset consisting of 257 cells belonging to five types of leukocytes was used. The results obtained from experiments show that the proposed algorithm attained a mean average precision of 98.42% (@IoU = 0.7) in white blood cell localization and a classification accuracy of 98.87 ± 1% compared to existing methods.