Label-Free Estimation of Therapeutic Efficacy on 3D Cancer Spheres Using Convolutional Neural Network Image Analysis.

Despite recent advances in cancer treatment, developing better therapeutic reagents remains an essential task for oncologists. To accurately characterize drug efficacy, 3D cell culture holds great promise as opposed to conventional 2D monolayer culture. Due to the advantages of cell manipulation in high-throughput, various microfluidic platforms have been developed for drug screening with 3D models. However, the dissemination of microfluidic technology is overall slow, and one missing part is fast and low-cost assay readout. In this work, we developed a microfluidic chip forming 1920 tumor spheres for drug testing, and the platform is supported by automatic image collection and cropping for analysis. Using conventional LIVE/DEAD staining as the ground truth of sphere viability, we trained a convolutional neural network to estimate sphere viability based on its bright-field image. The estimated sphere viability was highly correlated with the ground truth (R-value > 0.84). In this manner, we precisely estimated drug efficacy of three chemotherapy drugs, doxorubicin, oxaliplatin, and irinotecan. We also cross-validated the trained networks of doxorubicin and oxaliplatin and found common bright-field morphological features indicating sphere viability. The discovery suggests the potential to train a generic network using some representative drugs and apply it to many different drugs in large-scale screening. The bright-field estimation of sphere viability saves LIVE/DEAD staining reagent cost and fluorescence imaging time. More importantly, the presented method allows viability estimation in a label-free and nondestructive manner. In short, with image processing and machine learning, the presented method provides a fast, low-cost, and label-free method to assess tumor sphere viability for large-scale drug screening in microfluidics.