Unsupervised Cross-Domain White Blood Cells Classification Using DANN

The classification of white blood cells (WBCs) from microscopic blood image provides invaluable information for diagnosis of various diseases. Deep Convolutional Neural Networks are often used to classify WBCs automatically and have obtained certain achievements. However, when the training (source) dataset and test (target) dataset fall from different data distributions (i.e. domain shift), deep convolution neural networks adapt poorly. To solve the problem, we proposed a DANN-based method aiming to help our classifier learn domain-invariant information by using adversarial training. Two datasets were tested and our method achieved 97.1% accuracy, 97.2% recall, 97.2% precision and 97.4%f1-score, respectively. Domain adaptation verification shows that the proposed method has higher performance than other adaptive methods, and has broad application prospects in WBC classification.