The characterization and classification of white blood cells (WBC) are critical for the diagnosis of anemia, leukemia, and many other hematologic diseases. We developed WBC-Profiler, an unsupervised feature learning system for quantitative analysis of leukocytes. We demonstrate, through independent validation, that WBC-Profiler enables automatic extraction of complex and robust signatures from microscopic images without human-intervention and, thereafter, effective construction of interpretable leukocyte profiles, which decouples large scale complex leukocyte characterization from limitations in both human-based feature engineering/optimization and the end-to-end solutions provided by many modern deep neural networks. Further evaluation in a real-world clinical setting confirms that, compared with 23 clinicians from 8 hospitals (class-average-sensitivity, 0.798; class-average-specificity, 0.963; cell-average-timecost: 3.158 s), WBC-Profiler performs with significantly improved accuracy and speed (class-average-sensitivity, 0.890; class-average-specificity, 0.980; cell-average-timecost: 0.375 s). Our findings suggest that WBC-Profiler has the potential clinical implications.