Rapid cancer diagnosis using deep learning-powered label-free subcellular-resolution photoacoustic histology.

Abstract

Traditional hematoxylin and eosin staining in formalin-fixed paraffin-embedded sections, while essential for diagnostic pathology, is time-consuming, labor intensive, and prone to artifacts that can obscure critical histological details. Label-free ultraviolet photoacoustic microscopy (UV-PAM) has emerged as a promising alternative, offering fast histology-like images without the need for traditional staining and excessive tissue preparation. However, current UV-PAM systems face challenges in achieving the high spatial resolution required for detailed histological analysis and diagnosis. To address this, we developed a subcellular-resolution UV-PAM (SRUV-PAM) system with a 240-nanometer resolution, enabled by the integration of a high numerical aperture (NA) objective lens (NA = 0.64) and the precise piezo actuators for fine scanning control. This configuration allows visualization of detailed nuclear structures. In addition, we demonstrated virtual staining of SRUV-PAM images via cycle-consistent generative adversarial networks and diagnosis of malignant and benign tumors in liver tissues via densely connected convolutional networks DenseNet-121, achieving an area under the receiver operating characteristic curve of 0.902.