A Decision support system to identify downey cells from leukocytes via artificial intelligence

Abstract

BACKGROUND AND AIM: Downey cell is the atypical lymphocyte of infectious mononucleosis. It can be seen in a variety of conditions, but are often increased in infectious mononucleosis due to Epstein-Barr virus (EBV) infection. These cells can be identified through peripheral blood smear (PBS) for the diagnosis. Manual microscopy is the gold standard to evaluate Wright-Giemsa stained peripheral blood smear (PBS). As an alternative to manual microscopy, slide scanner systems (Mantiscope, Sysmex, etc…) are used to digitize the biological samples and make them available for physicians to analyze the samples with Artificial Intelligence (AI) tools. The main objective of this study is to create a decision support system (using AI) that can alert the physician for this type of cells to increase the rate of true diagnosis. METHODS: A dataset (PBS slides) examined via manual microscopy is collected from 25 patients with infectious mononucleosis and 124 healthy controls. Mantiscope slide scanner system is used to digitize the PBS slides. Each slide is scanned with 100x magnification via collecting approximately 400 images per slide. Physicians have annotated the digitized images through Mantiscope’s cloud-based annotation platform and labeled the cells as leukocytes, blasts and Downey cells. Data augmentation methods such as mirroring in vertical/ horizontal/ vertical and horizontal are used to create a balanced dataset to train deep learning-based AI platform. Objective measurement metrics (sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV)) are calculated to create insights about the effectivity of the decision support system. RESULTS: For a large variety of dataset comprising each normal leukocyte type and including blasts, identification of Downey can be performed with a higher sensitivity and specificity of 0.97 and 0.98, respectively. Cohen’s kappa analysis was conducted to evaluate the agreement between physicians and AI in terms of the presence of Downey cells. Kappa value was calculated as 0,93 and obtaining this value between 0,81 and 1,00 was accepted as an indicator of high agreement. CONCLUSIONS: This algorithm for the identification of Downey cells to facilitate the accurate diagnosis can be applied over the digitized images collected from the slide scanners and can also provide a helpful preliminary diagnosis to the physician. AI-based systems in hematology will help the physicians in the diagnosis and treatment of diseases and their success rates will increase day by day.