CellDiffuser: Multimodal Optical Time-Stretch Imaging Flow Cytometry and Diffusion Models for Lung Cancer Immunotherapy Diagnosis
TL;DR
This work establishes a cost-effective, high-throughput framework for single-cell phenotyping, offering a new paradigm to guide lung cancer personalized immunotherapy diagnosis.
Abstract
Immune checkpoint inhibitors have emerged as frontline therapies for lung cancer, yet the lack of reliable biomarkers for predicting the treatment response remains a critical challenge. While single-cell analysis offers a potent approach to probing physiological states, conventional techniques suffer from limited throughput and insufficient multimodal profiling capability. To address the aforementioned issues, we propose a novel framework for lung cancer immunotherapy diagnosis. Specifically, we develop multimodal optical time-stretch imaging flow cytometry (OTS-IFC) that simultaneously acquires label-free bright-field (BF) and quantitative phase (QP) images at a throughput of 10,000 cells/s. In response to the influence of phase unwrapping artifacts, we propose CellDiffuser, an intermediate domain transformation diffusion model that aligns BF and QP images in a shared latent space at a preselected time step, thus generating high-fidelity synthesized QP images. CellDiffuser achieves an SSIM of 0.747 in the leukocyte immunotherapy response prediction data set and an SSIM of 0.708 in the A549 cell death state identification data set. Finally, our framework demonstrates its practical utility in lung cancer immunotherapy diagnosis in the above two clinically relevant data sets. This work establishes a cost-effective, high-throughput framework for single-cell phenotyping, offering a new paradigm to guide lung cancer personalized immunotherapy diagnosis. The code is available at: https://github.com/yzygit1230/CellDiffuser.