Decision support system for the classification of Downey cells as a pre-diagnostic tool

Abstract

Abstract Objectives Epstein–Barr virus (EBV) is a member of the herpes virus that causes infectious mononucleosis (IM). Downey cell is the atypical lymphocyte of IM and can be seen in various conditions. Peripheral blood smear (PBS) microscopic evaluation is used to identify Downey cells. A lack of experienced professionals or professional errors may obstruct early and accurate diagnostics for the microscopic evaluation. The main objective of this study is to create a decision support system by digitizing the PBS samples. A general tool providing an inexpensive and measurable solution is envisioned to analyze the PBS samples in detail to give alerting flags to prevent missing Downey cells in manual analysis. Methods The PBS dataset collected was split into Downey positives and negatives. The negative set consisted of 5 leucocyte subtypes. Mantiscope, a cloud-based slide scanner system, was used to collect images from the physical PBS samples. Clinically and laboratory-confirmed 35 IM patients and 124 healthy PBS slides were selected for this procedure. A number of cell counts were obtained after the application of annotation and augmentation methods, and a partially balanced dataset was created for the artificial intelligence (AI) network training. The verification steps included the calculation of sensitivity, specificity, and Cohen’s kappa metrics from the partitioned testing set that was not used during training. A validation process was also performed over the manually identified PBS samples to measure whether the algorithm noticed the samples or not. Results After testing this setup, we have observed 98 % sensitivity and 99 % specificity for Downey cells. According to the validation procedure of Downey positive and negative samples that were carried out by the physicians, a sensitivity of 57 %, specificity of 100 %, and Cohen’s kappa value of 0.5 were observed. Besides, the accuracy was found to be 66 % according to the physicians’ evaluations employing the digital images which were identified by Mantiscope, Conclusions Decision support systems can alert the physician for Downey cells and increase the rate of true diagnosis in PBS evaluation. A higher sensitivity and specificity for the detection of Downey cells would be achieved. However, the variance over the dataset is a constraint for effective diagnosis. As the annotation and AI development process continues to collect more data from patients, the model can be updated for future releases.