Automating Pitted Red Blood Cell Counts Using Deep Neural Network Analysis: A New Method for Measuring Splenic Function in Sickle Cell Anaemia

Abstract

The spleen plays an important role in the body’s defence against bacterial infections. Measuring splenic function is of interest in multiple conditions, including sickle cell anaemia (SCA), where spleen injury occurs early in life. Unfortunately, there is no direct and simple way of measuring splenic function, and it is rarely assessed in clinical or research settings. Manual counts of pitted red blood cells (RBCs) observed with differential interference contrast (DIC) microscopy is a well-validated surrogate biomarker of splenic function. The method, however, is both user-dependent and laborious. In this study, we propose a new automated workflow for counting pitted RBCs using deep neural network analysis. Secondly, we assess the durability of fixed RBCs for pitted RBC counts over time. We included samples from 48 children with SCA and 10 healthy controls. Cells were fixed in paraformaldehyde and examined using an oil-immersion objective, and microscopy images were recorded with a DIC setup. Manual pitted RBC counts were performed by examining a minimum of 500 RBCs for pits, expressing the proportion of pitted RBCs as a percentage (%PIT). Automated pitted RBC counts were generated by first segmenting DIC images using a Zeiss Intellesis deep learning model, recognising and segmenting cells and pits from background. Subsequently, segmented images were analysed using a small ImageJ macro language script. Selected samples were stored for 24 months, and manual pitted RBC counts performed at various time points. When comparing manual and automated pitted RBC counts, we found the two methods to yield comparable results. Although variability between the measurements increased with higher %PIT, this did not change the diagnosis of asplenia. Furthermore, we found no significant changes in %PIT after storing samples for up to 24 months and under varying temperatures and light exposures. We have shown that automated pitted RBC counts, produced using deep neural network analysis, are comparable to manual counts, and that fixed samples can be stored for long periods of time without affecting the %PIT. Automating pitted RBC counts makes the method less time consuming and results comparable across laboratories.