Performance of the automated digital cell image analyzer UIMD PBIA in white blood cell classification: a comparative...

This study aimed to evaluate the performance of PBIA (UIMD, Seoul, Republic of Korea), an automated digital morphology analyzer using deep learning, for white blood cell (WBC) classification in peripheral blood smears and compare it with the widely used DI-60 (Sysmex, Kobe, Japan).

A total of 461 slides were analyzed using PBIA and DI-60. For each instrument, pre-classification performance was evaluated on the basis of post-classification results verified by users. Pre- and post-classification results were compared with manual WBC differentials, and the ability to identify abnormal cells was assessed.

The pre-classification performance of PBIA was better than that of DI-60 for most cell classes. PBIA had an accuracy of 90.0 % and Cohen's kappa of 0.934, higher than DI-60 (45.5 % accuracy and 0.629 kappa) across all cell classes. The pre-classification performance of both instruments decreased when abnormal cells were observed in manual counts, but PBIA still performed better. PBIA also appeared to show better correlation with manual WBC differential counts, particularly in pre-classification (Pearson's correlation coefficient: 0.696-0.944 vs. 0.230-0.882 for neutrophils, lymphocytes, monocytes, eosinophils, basophils, and blasts), although the mean differences varied by cell class. For abnormal cells identified in manual counts, PBIA exhibited more false positives for blasts (30.5 vs. 2.3 %), while DI-60 had a higher rate of false negatives (42.1 vs. 6.1 %). Both instruments exhibited high false negative rates for atypical lymphocytes.

PBIA demonstrated better performance than DI-60, highlighting its clinical utility. Further multicenter studies are required for full validation.