Deep learning-based adaptive detection of fetal nucleated red blood cells

Abstract

15Aim: this study, we established an artificial intelligence system for rapid 16 identification of fetal nucleated red blood cells (fNRBCs). 17Method: Density gradient centrifugation and magnetic-activated cell sorting were 18 used for the separation of fNRBCs from umbilical cord blood. The cell block 19 technique was used for fixation. We proposed a novel preprocessing method based on 20 imaging characteristics of fNRBCs for region of interest (ROI) extraction, which 21 automatically segmented individual cells in peripheral blood cell smears. The 22 discriminant information from ROIs was encoded into a feature vector and 23 pathological diagnosis were provided by the prediction network. 24Results: Four umbilical cord blood samples were collected and validated based on a 25 large dataset containing 260 samples. Finally, the dataset was classified into 3,720 and 26 1,040 slides for training and testing, respectively. In the test set, classifier obtained 27 98.5% accuracy and 96.5% sensitivity. 28Conclusion: Therefore, this study offers an effective and accurate method for 29 fNRBCs preservation and identification. 30 Keywords: fetal nucleated red blood cells, cell-block, deep learning, non-invasive 31 prenatal diagnosis 32 Introduction 33The clinical application of fNRBCs during pregnancy could be classified into two 34 main categories 1,2 . One is the prognosis of possible diseases in pregnant women by 35 counting fNRBCs in umbilical cord blood. Chronic tissue hypoxia results in increased 36 levels of erythropoietin, which, in turn, leads to stimulation of erythropoiesis and 37 increased numbers of circulating nucleated red blood cells (NRBCs) 1,3-5 . Increased 38 umbilical cord levels of erythropoietin have been reported in pregnancies complicated 39 by intrauterine growth restriction, maternal hypertension, preeclampsia, maternal 40 smoking, Rh isoimmunization, and maternal diabetes 5-8 . As expected, each of these 41 conditions has been associated with increased NRBCs in the newborn 9 . The other 42 objective is to screen and extract fNRBCs from maternal peripheral blood for 43 non-invasive prenatal diagnosis (NIPD) 10-12 . The choice of fNRBCs as ideal target 44 cells is based on the following parameters 13,14 : (1) presence of intact nuclei containing 45 the complete fetal genome in fNRBCs, which is a prerequisite for prenatal analysis; (2) 46 limited life span of fNRBCs in the maternal circulation, which can be differentiated 47 55 cells) 16,17 . Several fNRBC enrichment methods based on different principles have 56 been reported, such as density gradient centrifugation (DGC) 13,18 , 57 fluorescence-activated cell sorting (FACS) 19 , and magnetic-activated cell sorting 58 (MACS) 20 , dielectrophoresis, and microfluidics based technologies 21,22 . Nevertheless, 59 long-term preservation of samples and rapid identification of target cells (fNRBCs) 60 still present considerable challenges. 61 Since the identification of fNRBCs in large number of cell block slices represent a 62 huge manual burden on p...