Perovskite\nProbe-Based Machine Learning Imaging Model\nfor Rapid Pathologic Diagnosis of Cancers

Accurately\ndistinguishing tumor cells from normal cells is a key\nissue in tumor diagnosis, evaluation, and treatment. Fluorescence-based\nimmunohistochemistry as the standard method faces the inherent challenges\nof the heterogeneity of tumor cells and the lack of big data analysis\nof probing images. Here, we have demonstrated a machine learning-driven\nimaging method for rapid pathological diagnosis of five types of cancers\n(breast, colon, liver, lung, and stomach) using a perovskite nanocrystal\nprobe. After conducting the bioanalysis of survivin expression in\nfive different cancers, high-efficiency perovskite nanocrystal probes\nmodified with the survivin antibody can recognize the cancer tissue\nsection at the single cell level. The tumor to normal (T/N) ratio\nis 10.3-fold higher than that of a conventional fluorescent probe,\nwhich can successfully differentiate between tumors and adjacent normal\ntissues within 10 min. The features of the fluorescence intensity\nand pathological texture morphology have been extracted and analyzed\nfrom 1000 fluorescence images by machine learning. The final integrated\ndecision model makes the area under the receiver operating characteristic\ncurve (area under the curve) value of machine learning classification\nof breast, colon, liver, lung, and stomach above 90% while predicting\nthe tumor organ of 92% of positive patients. This method demonstrates\na high T/N ratio probe in the precise diagnosis of multiple cancers,\nwhich will be good for improving the accuracy of surgical resection\nand reducing cancer mortality.