Pancreatic cancer has a five-year survival rate below 11% despite treatment with aggressive chemotherapy regimens. A key reason drugs fail is intratumoral heterogeneity (ITH), the fact that tumors contain many different cell subtypes with different biological behaviors. When chemotherapy kills most cancer cells, the resistant subpopulation survives and repopulates the tumor.
Current methods to measure ITH, such as single-cell RNA sequencing, are expensive, slow, and technically complex. They cannot realistically be run in real time during patient care. There is an urgent need for faster, cheaper tools that can monitor how heterogeneous a tumor is and how it changes during treatment.
Digital holographic microscopy (DHM) uses laser light interference to generate detailed images of individual cells without any staining or labeling, which can take days in traditional methods. When light passes through a cell, it creates a phase image capturing the cell's thickness, dry mass, and internal structure. These physical properties differ between cell types and change with the cell's biological state.
The researchers coupled a DHM device to a microfluidic system so that cancer cells in liquid suspension flow past the microscope in a continuous stream. This enables high-throughput single-cell imaging at speeds not possible with traditional microscopy, collecting data on thousands of individual cells in minutes rather than hours or days.
The team built a machine learning pipeline combining traditional morphological features (cell shape, area, perimeter) with features extracted by a lightweight neural network called ResNet18. A random forest classifier, tested against support vector machines, K-nearest neighbors, and neural networks, performed best throughout all experiments.
The model was first validated in spike-in experiments, mixing known proportions of two distinct cell lines (epithelial and mesenchymal) and testing whether the AI could recover the correct mixture ratios. It achieved this with a maximum deviation of only 3%, confirming the system's precision before applying it to more complex biological questions.
When pancreatic cancer cells were treated with TGF-beta to trigger epithelial-mesenchymal transition (a key step toward cancer spread), the DHM system detected the resulting cell shape changes with 81-85% accuracy in two of three cell lines. The system also correctly identified that one cell line (8442) showed minimal response, consistent with immunofluorescence validation experiments.
Most importantly, the system was applied to patient-derived organoids, three-dimensional tumor structures grown from actual patient tissue samples. The DHM pipeline successfully detected heterogeneity within these organoids and identified changes in cell state in response to gemcitabine chemotherapy treatment, demonstrating the technology's translational potential with clinically relevant models.
This platform could be integrated into clinical workflows to provide rapid phenotypic profiling of tumor biopsies or cells obtained from liquid biopsies. By monitoring how heterogeneous a patient's tumor is and how it evolves under treatment, oncologists could make more informed decisions about when to switch chemotherapy regimens or add targeted therapies.
The label-free, real-time nature of DHM means results can theoretically be available within hours of a biopsy rather than weeks. As the technology matures, it could become a routine tool for personalizing pancreatic cancer treatment, helping identify which patients are developing resistant cell populations before clinical relapse occurs.