Interpretable Multi-Task Conditional Neural Networks Reveal Cancer Cell Adhesion Characteristics From Phonon Microscopy Images.

Abstract

Advances in artificial intelligence (AI) show significant promise in multiscale modeling and biomedical informatics, particularly in the analysis of phonon microscopy (high-frequency ultrasound) data for cancer detection. This study addresses critical issues in data engineering for time-resolved phonon microscopy of biomedical samples by tackling the 'batch effect,' which arises from unavoidable technical variations between experiments, creating confounding variables that AI models may inadvertently learn. We present a multi-task conditional neural network framework that simultaneously achieves inter-batch calibration by removing confounding variables and accurate cell classification from time-resolved phonon-derived signals. We validate our approach by training and validating on different experimental batches, achieving a balanced precision of 89.22% and an average cross-validated precision of 89.07% for classifying background, healthy and cancerous regions. Furthermore, our model enables reconstruction of denoised images, which enable the physical interpretation of salient features indicative of disease states, such as sound velocity, sound attenuation, and cell adhesion to substrates. This work demonstrates the potential of AI methodologies in improving health outcomes and advancing cancer-informatics platforms.