Classifying Microscopic Images of Reactive Lymphocytosis Using Two-Step Tandem AI Models

The practical applications of automatic recognition and categorization technology for next-generation systems are desired in the clinical laboratory. We approached the identification of reactive lymphocytosis using artificial intelligence (AI) technology and studied its clinical usefulness for blood smear screening. This study created one- and two-step AI models for the identification of reactive lymphocytosis. The ResNet-101 model was applied for deep learning. The original image set for supervised AI training consisted of 5765 typical nucleated blood cell images. The subjects for clinical assessment were 25 healthy cases, 25 erythroblast cases, and 25 reactive lymphocytosis cases. The total accuracy (mean ± standard deviation) of the one- and two-step models were 0.971 ± 0.047 and 0.977 ± 0.024 in healthy, 0.938 ± 0.040 and 0.978 ± 0.018 in erythroblast, and 0.856 ± 0.056 and 0.863 ± 0.069 in reactive lymphocytosis cases, respectively. The two-step AI model showed a sensitivity of 0.960 and a specificity of 1.000 between healthy and reactive lymphocytosis cases. As our two-step tandem AI model showed high performance for identifying reactive lymphocytosis in blood smear screening, we plan to apply this method to the development of AI models to differentiate reactive and neoplastic lymphocytosis.