Accurate tumor cell identification via AI analysis of white light scattering spectra in cytology

Detecting subtle intracellular structural alterations is important for improving diagnostic accuracy. Although such structures are often difficult to evaluate directly by conventional bright-field microscopic observation, they influence visible light scattering. In this study, we propose a diagnostic method utilizing the light scattering that originates Rayleigh and Mie scattering due to the fine structures. We measured the scattering spectrum of cells in cytology specimens under dark-field illumination, i.e. the white-light scattering spectrum, and analyzed them using machine learning classification. We used Papanicolaou stained cytology specimens of pleural effusion. The spectra of these individual cells were analyzed by principal component analysis to extract the principal components (PCs) including the spectral features. The identification of mesothelioma cells was performed by a support vector machine classification, using all PCs to distinguish the cells. The diagnostic accuracy, which is the probability of diagnosed mesothelioma cells being among the tested mesothelioma cells, was 99 ± 1% and 92 ± 10% in the five-fold cross-validation with random groupings and in the validation with patient-based groupings, respectively. Similar results were obtained for gastric adenocarcinoma in ascites and urothelial carcinoma in urine. Furthermore, lung adenocarcinoma was distinguished from pleural mesothelioma, small cell lung cancer and gastric adenocarcinoma with over 80% accuracy. These high cell identification rates are presumed to be based on intracellular submicron structures with sizes comparable to and smaller than the light wavelength. Integrating this spectroscopic system with a microscope could serve as a powerful aid for pathologists and potentially improve the accuracy of their diagnosis.