Background: The rapid recognition of fetal nucleated red blood cells (fNRBCs) present considerable challenges.
Objective: To establish a computer-aided diagnosis system (CAD) for rapid recognition of fNRBCs by a convolutional neural network (CNN).
Methods: We adopted density gradient centrifugation and magnetic-activated cell sorting to extract fNRBCs from umbilical cord blood samples. A cell-block method was used to embed fNRBCs for routine formalin-fixed paraffin sectioning and hematoxylin-eosin stains. Then we proposed a CAD-based on CNN to automatically learn discriminative features and recognize fNRBCs. Region of interest 1 extraction methods were used to automatically segment individual cells in cell slices. The discriminant information from ROIs was encoded into a feature vector. The prediction network provided a pathological diagnosis.
Results: Totally, 4760 pictures of fNRBCs from 260 cell-slides of 4 umbilical cord blood samples were collected. On the premise of 100% accuracy in the training set (3720 pictures) the sensitivity, specificity, and accuracy of cellular intelligent recognition were 96.5%, 100%, and 98.5% in the test set (1040 pictures).
Conclusion: We present a CAD system for effective and accurate fNRBCs recognition based on CNN.