A Convolutional Neural Network- Based Deep Learning To Detect Reticulocytes From Human Peripheral Blood

Machine learning approaches are rapidly augmenting, and in some cases, replacing the conventional methods in biomedical data analysis; to reduce time, cost, biases, and the need for sophisticated analytical platforms. Hence, significant interest has been compounded in the integration of automated image analysis for various clinical applications, such as the detection of infected or inflamed wounds, bone fractures or for the purpose of disease diagnosis – such as Plasmodium parasites or circulating tumour cells in blood. Here, we report the development of a Convolutional Neural Network (CNN)-based method on CPU to distinguish and count immature human red blood cells known as reticulocytes from blood smears. Reticulocytes represent a heterogeneous and relatively small percentage of cells in peripheral blood, and contain residual RNA in complex with proteins which generates thread-like patterns when stained with New Methylene Blue (NMB) dye. We used more than 200 NMB-stained images from leukocyte-depleted blood to train and optimize the model for immature reticulocytes (stained positive with NMB, intensity and pattern of which depends on the developmental stage of the reticulocyte) and mature RBCs (no staining with NMB). The training performance evaluation metrics demonstrated a mean average precision (mAP50) of 0.88, a precision of 0.83, a recall of 0.88, and an F1 score of 0.87. Our model was able to successfully count reticulocytes with accuracy more than 90% from unknown samples which were subsequently cross-verified through microscopy and counting. Given the importance of reticulocyte maturation and its clinical relevance, the newly developed model will find important, easy to adopt biomedical applications that can be achieved on a simple PC.