A 3D attention networks for classification of white blood cells from microscopy hyperspectral images

Peripheral blood smear analysis is a key process for hematologists to reflect condition of human immune system and one of the earliest clinical applications that benefit from automatic computer-aided analysis. Changes in the ratio of white blood cell (WBC) types are associated with the determination of blood diseases, and therefore accurate classification assures accurate therapy. In this study, we applied deep convolution networks on microscopy hyperspectral images for WBC classification. We proposed a 3D convolutional networks named deep hyper that enables learning spectral and spatial features by itself to make fully use of three dimensional hyperspectral data for white blood cell classification. In order to enhance the representative power of the model in an efficient manner, we integrated a 3D attention module with the last block of the model to put more emphasis on important features. The overall classification accuracy of deep hyper can achieve 96% compared to 90% of the machine learning based method and the attention module allows deep hyper to achieve the best performance of 97.72%. Furthermore, we explored the correlation between spectral features and WBC classification performances to present that hyperspectral characteristics play an important role in classifying specific type of WBC. These findings demonstrate that combing microscopy hyperspectral image with deep convolution networks is beneficial for blood smear analysis especially white blood cell classification.