Abstract
Accurate cancer detection and diagnosis is of utmost importance for reliable drug-response prediction. Successful cancer characterization relies on both genetic analysis and histological scans from tumor biopsies. It is known that the cytoskeleton is significantly altered in cancer, as cellular structure dynamically remodels to promote proliferation, migration, and metastasis. We exploited these structural differences with supervised feature extraction methods to introduce an algorithm that could distinguish cancer from non-cancer cells presented in high-resolution, single cell images. In this paper, we successfully identified the features with the most discriminatory power to successfully predict cell type with as few as 100 cells per cell line. This trait overcomes a key barrier of machine learning methodologies: insufficient data. Furthermore, normalizing cell shape via microcontact printing on self-assembled monolayers enabled better discrimination of cell lines with difficult-to-distinguish phenotypes. Classification accuracy remained robust as we tested dissimilar cell lines across various tissue origins, which supports the generalizability of our algorithm.