Phase-augmented deep learning for cell segmentation in wrapped quantitative phase images.

Understanding cell adhesion and detachment is crucial for advancing disease diagnosis, treatment, and biomaterial development. Optical phase imaging techniques enable continuous, label-free observation of cells undergoing dynamic processes, including cell adhesion and detachment. To quantitatively study these processes with single-cell precision, accurate cell segmentation from phase images is essential. However, it is challenging to perform cell segmentation in phase images because of phase-wrapping artifacts. In this study, we developed a phase-augmented deep learning approach for cell segmentation in wrapped quantitative phase images. We acquired phase images from modulated optically computer phase microscopy (M-OCPM), a novel imaging technology recently developed in our laboratory. We trained a neural network with U-Net architecture to perform cell segmentation. The novelty of our method lies in the data augmentation strategy that introduces global phase shifts to the input images, enabling the network to distinguish true morphological features from phase-wrapping artifacts. It results in improved segmentation accuracy and eliminates the need for unwrapping. With the phase-augmented U-Net segmentation, we performed a quantitative analysis of cell morphology during cell detachment, highlighting the value of deep learning segmentation for studying dynamic cellular processes.