The utilization of the k-means clustering for cancer cell detection and classification with serous effusion.

Abstract

Cytological analysis of serous effusion specimens is essential for cancer diagnosis. In this work, we analyzed three-dimensional (3D) morphologic features by clustering to discriminate between malignant and nonmalignant cells in serous effusion specimens collected from 10 patients with pleural and peritoneal effusion accumulation symptoms. After the nuclei and mitochondria were fluorescently labeled, we obtained confocal image stack data and conducted 3D reconstruction and morphological feature parameter computation. Confocal images were segmented, interpolated, and reconstructed. Quantitative comparison across cell types has been made by 27 morphological features of volume and surface linked to the cell, nucleus, and mitochondria. We used an unsupervised machine learning method ofk-meansclustering to separate the cell distribution objectively and effectively in the 3D parameter space of the cell morphology features. The statistical significance of the differences was examined on morphological features among the three cell clusters. The clustering results were also analyzed against those of cytopathological examinations performed by collaborative pathologist on specimens collected from the same patients. These results showed that 3D morphologic features allow clustering of the effusion cells in the space of these parameters and may help produce new ways to quickly profile cells for cancer diagnosis in clinical settings. By incorporating these techniques into clinical practice, healthcare professionals may be able to more effectively detect and treat cancers in patients.