The traditional diagnosis of leukemia relies on pathologists to observe and classify cells on bone marrow smears, which is low-throughput, time-consuming, and subject to human bias. To overcome these limitations, we demonstrate intelligent frequency-shifted optofluidic time-stretch quantitative phase imaging (OTS-QPI) that acquires bright-field and quantitative phase images of white blood cells (WBCs) containing leukemia cells with high throughput (15,000 cells/s) for deep-learning-based classification. After trained with 64,000 images, a convolutional neural network (CNN) distinguishes three different types of leukemia cells from WBCs with an accuracy of over 96%. Our method provides new possibilities for high-throughput, label-free, and intelligent leukemia diagnosis.
|