Leveraging Pre-trained Neural Network Models for the Classification of Tumor Cells Analyzed by Label-free Phase Holotomographic Microscopy

Abstract

Can a single label-free image reveal whether cancer cells were exposed to chemotherapy? We present an innovative methodology on the label-free and high-resolution imaging properties of phase holotomographic microscopy coupled with neural network models for the classification of cancer cells. Using 3D phase holotomographic microscopy, we imaged live A549 lung cancer cells with and without paclitaxel, converted stacks to 2D maximum-intensity projections, and evaluated pre-trained convolutional networks (VGG16, ResNet18, DenseNet121, and EfficientNet-B0) for binary classification of treatment status. EfficientNet-B0 achieved 96.9 % accuracy on unsegmented images. Refractive index analysis revealed bimodal distribution in treated cells, reflecting heterogeneous biophysical responses to paclitaxel exposure and supporting the network's ability to detect subtle, label-free indicators of drug action. As further proof-of-concept, the same pipeline separated holotomographic images of label-free, high versus low-graded urothelial cancer cells with high accuracy (90.6 %). These findings highlight the potential of integrating label-free holotomographic imaging with deep learning techniques for rapid and efficient classification of tumor cells, paving the way for advancements in treatment optimization and personalized diagnostic strategies.