Real-Time, Label-Free Classification of Cell Death Pathways via Holotomography-Based Deep Learning Framework

Prof. YongKeun Park’s group in the Department of Physics at KAIST, in collaboration with Prof. Won Do Heo’s group, has developed a real-time, label-free platform for classifying cell death pathways by integrating 3D holotomography (HT) with deep learning. The platform identifies five distinct cell states—apoptosis, necroptosis, necrosis, untreated live, and drug-treated live—directly from volumetric refractive index (RI) maps, without any fluorescent labeling, fixation, or staining. The study was published online in Advanced Intelligent Systems (“Real-Time, Label-Free Classification of Cell Death Pathways via Holotomography-Based Deep Learning Framework,” Advanced Intelligent Systems, DOI: 10.1002/aisy.202500633).

Conventional fluorescence microscopy—the gold standard for cell death classification—requires invasive labeling procedures that limit live-cell imaging duration and introduce phototoxicity. Existing label-free methods such as 2D quantitative phase imaging (QPI) lack the volumetric depth resolution needed to distinguish morphologically similar death pathways. The new framework addresses these limitations by exploiting 3D HT, which reconstructs high-resolution volumetric RI tomograms from multi-angle holographic measurements using the Tomocube HT-X1 instrument. Each 3D RI volume is converted into a 2D maximum intensity projection (MIP) and subdivided into fixed-size patches (160 × 160 pixels; 25.6 µm per side) for inference by a fine-tuned ResNet-101 convolutional neural network. Ground-truth annotations for model training were generated by co-registering RI and fluorescence images (Hoechst, Annexin V, PI) of chemically treated HeLa cells.

The trained model achieved an overall accuracy of 99.3% on an independent test set, with all five classes showing near-perfect area under the ROC curve (AUC = 1.00). Critically, the model detected early RI changes characteristic of necroptosis 2–4 hours before the appearance of conventional fluorescence markers such as Annexin V and propidium iodide (PI)—a finding validated against population-level flow cytometry data. The framework was further demonstrated on wide-field stitched images covering approximately 150 HeLa cells under densely populated conditions (up to 70% confluency), maintaining robust classification without requiring explicit single-cell segmentation. Adaptability to a distinct cell line (A549 lung cancer cells) was achieved through targeted fine-tuning, restoring near-perfect accuracy after an initial domain-shift performance drop. For complex drug-induced phenotypes (Doxorubicin and Cisplatin treatments), the full 3D volumetric model achieved 76–88% accuracy, while 2D projection methods failed entirely (0%), underscoring the indispensability of volumetric information for subtle phenotype discrimination.

“”Our holotomography-based AI platform can detect morphological changes associated with cell death earlier and more accurately than conventional fluorescence assays, all without any staining or labeling,” said Prof. YongKeun Park, the corresponding author. “This work establishes a scalable, label-free foundation for high-throughput drug screening, cytotoxicity profiling, and translational cell-fate analysis across diverse cell types and clinical contexts.”

Funding information was not explicitly stated in the paper. Y.K.P., S.O., J.D., and J.P. have financial interests in Tomocube, Inc., which commercializes the HT-X1 instrument used in this study.