Online Program

The Shape of Cancer: Identifying Morphologies of Precancerous Cells (304980)

Keywords: classification, machine learning, cancer, biostatistics, image processing

Cancer is an accumulation of rare, random, and spontaneous mutations, which makes detecting and studying pre-malignancy difficult. Modern microscopy enables collecting videos and extracting molecular features from individual cells grown in the lab. We analyze such single-cell-level data derived from patients with and without a genetic predisposition to cancer. Using image processing techniques, we identify cells in photographs and extract shape characteristics like “solidity,” a geometric measure of cell boundary smoothness. We use classification techniques, including LASSO, elastic net, SVMs, and random forests, to identify what features of cell shape consistently appear in cells with the pre-cancerous mutation. In particular, we benchmark our performance against deep learning methods with surprising results. To test the biological validity of our discoveries, we demonstrate the predictive power of these cell shape features in diagnosing malignancy. Future work will correlate cellular features with genetic features to discover biological pathways relevant to cancer development.