Background: The purpose of the present work was to test whether quantitative image analysis of circulating cells can provide useful clinical information targeting bone metastasis (BM) and overall survival (OS >30 months) in metastatic breast cancer (MBC).
Methods: Starting from cell images of epithelial circulating tumor cells (eCTC) and leukocytes (CD45pos) obtained with DEPArray, we identified the most significant features and applied single-variable and multi-variable methods, screening all combinations of four machine-learning approaches (Naïve Bayes, Logistic regression, Decision Trees, Random Forest).
Results: Best predictive features were circularity (OS) and diameter (BM), in both eCTC and CD45pos. Median difference in OS was 15 vs. 43 (months), p = 0.03 for eCTC and 19 vs. 36, p = 0.16 for CD45pos. Prediction for BM showed low accuracy (64%, 53%) but strong positive predictive value PPV (79%, 91%) for eCTC and CD45, respectively. Best machine learning model was Naïve Bayes, showing 46 vs 11 (months), p <0.0001 for eCTC; 12.5 vs. 45, p = 0.0004 for CD45pos and 11 vs. 45, p = 0.0003 for eCTC + CD45pos. BM prediction reached 91% accuracy with eCTC, 84% with CD45pos and 91% with combined model.
Conclusions: Quantitative image analysis and machine learning models were effective methods to predict survival and metastatic pattern, with both eCTC and CD45pos containing significant and complementary information.
Keywords: circulating tumor cells; data science; image analysis; liquid biopsy; machine learning.
Copyright © 2022 Da Col, Del Ben, Bulfoni, Turetta, Gerratana, Bertozzi, Beltrami and Cesselli.