Abstract

Doppler infrared spectroscopy of intracellular dynamics in living tumor tissue detects speeds down to nanometers per second (10 mHz) and up to microns per second (10 Hz) associated with a full range of cellular processes. Changes in these dynamics have specific Doppler signatures that depend on the applied cancer drugs and the sensitivity of the patient to treatment. However, strong intra-tumor heterogeneity poses a significant challenge to machine-learning classifiers. Here, we describe a Twin Deep Network (TDN) that can be trained to identify these signatures in the presence of strong heterogeneous background to accurately predict patient response to therapy. The TDN is applied to a clinical trial of HER2neg breast-cancer patients undergoing neoadjuvant therapy. This work provides insight into the value of Deep Learning for advanced data analytics as the volume and variety of data from optics-based assays grows.

© 2021 The Author(s)