White Blood Cells Recognition System Based on Deep Residual Network

摘要

Recently, several automated blood analysis systems (ABAS) have been developed based on image processing and artificial intelligence techniques. White blood cells differential counting and recognition aim to diagnose several human diseases. In the present study, a recognition system of five healthy white blood cells has been demonstrated. Deep learning methodologies have gained significant importance among artificial intelligence techniques for several ABAS. In this paper, we propose a white blood cells recognition system based on the residual deep network. We increase the performance of the previous recognition system based on three approaches which are residual blocks, dropout layer, and batch normalization layers. To save the computational power and cost time, we optimize the location of residual block inside the plain CNN network. A significant improvement has been achieved by the proposed system. The achieved results contribute to increasing either the traditional recognition system or deep learning system performance. During our experiments, we employ two white blood cells datasets which contain 2426 cropped images for different main five white blood cells (WBCs) classes. The proposed network achieves 96.8% accuracy and 96.4% sensitivity. Moreover, we re-employ the proposed system as a pre-trained network to recognize limited size WBCs dataset. Based on transfer learning, the proposed system achieved 95.83% for recognition limited size dataset. We visualize deep features to prove the power of our propped deep network. The experimental results show promising results for our proposed approach.