Atomic force microscopy-based assessment of multimechanical cellular properties for classification of graded bladder cancer cells and cancer early diagnosis using machine learning analysis

Cellular mechanical properties (CMPs) have been frequently reported as biomarkers for cell cancerization to assist objective cytology, compared to the current subjective method dependent on cytomorphology. However, single or dual CMPs cannot always successfully distinguish every kind of malignant cell from its benign counterpart. In this work, we extract 4 CMPs of four different graded bladder cancer (BC) cell lines by AFM (atomic force microscopy)-based nanoindentation to generate a CMP database, which is used to train a cancerization-grade classifier by machine learning. The classifier is tested on 4 categories of BC cells at different cancer grades. The classification shows split-independent robustness and an accuracy of 91.25% with an AUC-ROC (ROC stands for receiver operating characteristic, and ROC curve is a graph-ical plot which illustrates the performance of a binary classifier system as its discrimination threshold is varied) value of 97.98%. Finally, we also compare our proposed method with traditional invasive diag-nosis and noninvasive cancer diagnosis relying on cytomorphology, in terms of accuracy, sensitivity and specificity. Unlike former studies focusing on the discrimination between normal and cancerous cells, our study fulfills the classification of 4 graded cell lines at different cancerization stages, and thus provides a potential method for early detection of cancerization.Statement of significanceWe measured four cellular mechanical properties (CMPs) of 4 graded bladder cancer (BC) cell lines us-ing AFM (atomic force microscopy). We found that single or dual CMPs cannot fulfill the task of BC cell classification. Instead, we employ MLA (Machine Learning Algorithm)-based analysis whose inputs are BC CMPs. Compared with traditional cytomorphology-based prognoses, the non-invasive method pro-posed in this study has higher accuracy but with many fewer cellular properties as inputs. The proposed non-invasive prognosis is characterized with high sensitivity and specificity, and thus provides a poten-tial tumor-grading means to identify cancer cells with different metastatic potential. Moreover, our study proposes an objective grading method based on quantitative characteristics desirable for avoiding misdi-agnosis induced by ambiguous subjectivity.(c) 2022 Published by Elsevier Ltd on behalf of Acta Materialia Inc.

Automated summary

We measured the cellular mechanical properties (CMPs) of different graded bladder cancer (BC) cell lines using atomic force microscopy (AFM) and built a CMP database. Machine learning was employed to train a cancerization-grade classifier for classification of BC cells at different grades. Compared with traditional diagnostic methods, the proposed non-invasive prognosis has higher accuracy with fewer cellular properties as inputs. It provides a potential means for early detection of cancerization with high sensitivity and specificity.