Understanding the tumor microenvironment and detecting rare circulating tumor cells from blood are two major challenges faced by cancer biologists and oncologists. Both require high sensitivity and accuracy in classifying and isolating single cells in a complex and heterogeneous environment. Classical cell classification and sorting techniques are limited by their reliance on pre-selected cell biomarkers or physical characteristics. Recent breakthroughs in machine learning have achieved unprecedented accuracy across a wide range of image classification problems. We have developed a platform that combines high-resolution imaging of unlabeled cells in microfluidic flow with real-time deep neural network (DNN) based classification and sorting. The DNN classifier was trained on more than 25 million high-resolution cell images of multiple types imaged on the platform. Our model was trained to discriminate among multiple cell classes, including immune cell subtypes, non-small-cell lung cancer cells (NSCLC), hepatocellular carcinomas (HCC), and stromal cells (including endothelial, epithelial, fibroblasts, smooth muscle cells). We then assessed model performance on a separate validation set of cell images, including cell lines not used in the training data. Our classifier accurately identifies NSCLC and HCC against a background of blood cells with an area under the ROC curve (AUC) of > 0.999. In addition we demonstrate the enrichment of NSCLC cells from spike-in mixtures with WBCs or whole blood at concentrations as low as 1:100,000, achieving an enrichment of > 25,000x on multiple cell lines. Using dissociated lung cancer tissue, we demonstrate that our label-free classification of tumor cells closely matches results from both standard flow cytometer analysis and single cell RNA sequencing. Additionally we were able to enrich the tumor cell fraction from dissociated tumor tissue thereby improving the sensitivity of mutation detection and enabling refined downstream single-cell genomic analysis. This work demonstrates that deep learning using high-resolution cell images collected at scale can achieve a high classification accuracy and can enable the label-free isolation of rare cells of interest for a wide range of applications. This system can be used to analyze tumor biopsies and liquid biopsies and has the potential to enable the study of tumor cells and tumor microenvironment with novel dimension and insight. Citation Format: Mahyar Salek, Hou-pu Chou, Prashast Khandelwal, Krishna P. Pant, Thomas J. Musci, Nianzhen Li, Christina Chang, Andreja Jovic, Esther Lee, Stephanie Huang, Jeff Walker, Phuc Nguyen, Kiran Saini, Jeanette Mei, Quillan F. Smith, Maddison Masaeli. Deep learning enables label-free profiling of the tumor microenvironment and enrichment of rare cancer cells [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2021; 2021 Apr 10-15 and May 17-21. Philadelphia (PA): AACR; Cancer Res 2021;81(13_Suppl):Abstract nr 188.

No abstract