Deep learning approach to peripheral leukocyte recognition.

摘要: Microscopic examination of peripheral blood plays an important role in the field diagnosis and control major diseases. Peripheral leukocyte recognition by manual requires medical technicians to observe smears through light microscopy, using their experience expertise discriminate analyze different cells, which is time-consuming, labor-intensive subjective. The traditional systems based on feature engineering often need ensure successful segmentation then manually extract certain quantitative qualitative features for but still remaining a limitation poor robustness. classification pipeline convolutional neural network automatic extraction free hard deal with multiple object recognition. In this paper, we take as detection task apply two remarkable approaches, Single Shot Multibox Detector An Incremental Improvement Version You Only Look Once. To improve performance, some key factors involving these approaches are explored models generated train set 14,700 annotated images. Finally, evaluate test sets consisting 1,120 images 7,868 labeled single corresponding 11 categories leukocytes, respectively. A best mean average precision 93.10% accuracy 90.09% achieved while inference time 53 ms per image NVIDIA GTX1080Ti GPU.