High-accuracy and high-throughput reactive lymphocyte identification using lightweight neural networks

• Published in: Biomedical signal processing and control Vol. 97; p. 106722
• Main Authors: Mei, Liye, Jin, Shuangtong, Huang, Tingting, Peng, Haorang, Zha, Wenqi, He, Jing, Zhang, Songsong, Xu, Chuan, Yang, Wei, Shen, Hui, Lei, Cheng, Xiong, Bei
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.11.2024
• Subjects: Convolutional neural network; Deep learning; Lightweight; Peripheral blood; Reactive lymphocytes; Deep learning; Reactive lymphocytes; Lightweight; Convolutional neural network; Peripheral blood
• ISSN: 1746-8094
• DOI: 10.1016/j.bspc.2024.106722

•We propose an automatic method to identify reactive lymphocytes and enhance the diagnosis rate of infectious diseases.•We combine CNN and Transformer to extract complex cell features precisely.•We employ knowledge distillation to enhance the model’s discriminative ability.•We adopt a new peripheral blood cell dataset to assist in the diagnosis of infectious diseases. Infectious diseases caused by viruses are generally self-limited diseases with good prognoses, but their clinical manifestations and laboratory test results have a wide range of diversity. However, current clinical examination methods mainly focus on tumor diseases, which may lead to misdiagnosis and missed diagnosis of infectious diseases. Recognizing that the elevation of reactive lymphocytes is a significant indicator for diagnosing infectious diseases and considering the clinical necessity of integrating different cell proportion changes as a foundation for treatment, we have developed an effective and lightweight multi-classification model for diagnosing infectious diseases. Due to the diversity and difficulty of distinguishing reactive lymphocytes, we combine the efficiency of convolutional neural networks to model local features with the capability of the transformer to dynamically model remote interactions, which can extract complex features of cells while performing efficient processing operations. In addition, to further enhance the model’s recognition capabilities, we employ knowledge distillation techniques for data augmentation. According to our results, our model can identify six types of cell lines including reactive lymphocytes in peripheral blood with an average accuracy of 93.55% and can process more than 1300 cells/s per second, which greatly reduces labor costs. Providing early diagnosis of diseases and interventions for patients has significant practical implications. Artificial intelligence-based methods for morphometry of peripheral blood cells are further developed in this study, with potential clinical applications.