Leukocyte subtypes identification using bilinear self-attention convolutional neural network

• Published in: Measurement : journal of the International Measurement Confederation Vol. 173; p. 108643
• Main Authors: Yang, Dongxu, Zhao, Hongdong, Han, Tiecheng, Kang, Qing, Ma, Juncheng, Lu, Haiyan
• Format: Journal Article
• Language: English
• Published: London Elsevier Ltd 01.03.2021; Elsevier Science Ltd
• Subjects: Accuracy; Artificial neural networks; Bilinear strategy; Complexity; Convolutional neural network; Feature maps; Identification; Image segmentation; Leukocyte subtypes identification; Leukocytes; Medical imaging; Neural networks; Self-Attention mechanism; Tracking control systems; Visualization; Visualization; Bilinear strategy; Convolutional neural network; Self-Attention mechanism; Leukocyte subtypes identification
• ISSN: 0263-2241; 1873-412X
• DOI: 10.1016/j.measurement.2020.108643

•This work proposes a computational method for automatic classification of leukocyte subtypes.•This method uses a publicly available dataset of annotated blood cell images called Blood Cell Count Dataset (BCCD).•This novel lightweight CNN structure combines attention mechanisms and bilinear strategy.•The accuracy of this method on BCCD is 92.56%, ROC curve area is 0.95 and F1 score is 92.5%. Effective identification of leukocyte subtypes in microscopic images can help doctors diagnose diseases more accurately. Previous studies have achieved well performance by using segmentation techniques for multi-step processing. However, this increases the complexity of the whole identification process. In this paper, we proposed a novel model structure that can be trained end-to-end. The model combines attention mechanisms to emphasize the most discriminative features, and bilinear strategy to capture the interactions between features. We called this model Bilinear Self-Attention Network (BSA-Net). BSA-Net directly performs leukocyte subtypes identification in a one-step manner, which not only reduces complexity, but also achieves higher accuracy. To better understand the impact of the attention mechanism, we visualized the attention feature map in the BSA-Net model. Experiments demonstrated the effectiveness of our proposed method, which can meet the requirements of doctors for the accuracy and timeliness of cell identification results.