Label-free virtual peritoneal lavage cytology via deep-learning-assisted single-color stimulated Raman scattering microscopy

Abstract

Clinical guidelines for gastric cancer treatment recommend intraoperative peritoneal lavage cytology to detect free cancer cells. Patients with positive cytology require neoadjuvant chemotherapy instead of instant resection and conversion to negative cytology results in improved survival. However, the accuracy of cytological diagnosis by pathologists or artificial intelligence is disturbed by manually-produced, unstandardized slides. In addition, the elaborate infrastructure makes cytology accessible to a limited number of medical institutes. Here, we developed CellGAN, a deep learning method that enables label-free virtual peritoneal lavage cytology by producing virtual hematoxylin-eosin-stained images with single-color stimulated Raman scattering microscopy. A structural similarity loss was introduced to overcome the challenge of existing unsupervised virtual pathology techniques unable to present cellular structures accurately. This method achieved a structural similarity of 0.820±0.041 and a nucleus area consistency of 0.698±0.102, indicating the staining fidelity outperforming the state-of-the-art method. Diagnosis using virtually stained cells reached 93.8% accuracy and substantial consistency with conventional staining. Single-cell detection and classification on virtual slides achieved a mean average precision of 0.924 and an area under the receiver operating characteristic curve of 0.906, respectively. Collectively, this method achieves standardized and accurate virtual peritoneal lavage cytology and holds great potential for clinical translation.