Automated Focus Restoration for High-throughput Phase Contrast Time-lapse Microscopy with De-noising Diffusion Probabilistic Model

Cell-based immunotherapies have revolutionized cancer treatment with unprecedented efficacy against leukemias and lymphomas. High-throughput (HT) time-lapse imaging serves as a fundamental technology to enable the profiling of immune cells as anti-cancer drugs and to help map their complexity and heterogeneity. Migrating to label-free phase-contrast video microscopy promises non-invasive, non-toxic, and efficient profiling at single-cell level resolution. Balancing the tradeoffs between speed and throughput implies that despite the best autofocusing algorithms of HT systems, out-of-focus (OOF) cells are unavoidable due to the migratory nature of immune cells (velocities >10 μm/min). Restoration of OOF cells after acquisition increases the yield and accuracy of usable data by allowing more accurate segmentation and delineation of cell-cell contact.

Although traditional image restoration methods estimate the point spread function (PSF) for deconvolution and have made remarkable progress, is the approach tends to resource-intensive and time-consuming. Recent deep-learning-based approaches model the inverse of the PSF for restoring images. However, these methods generate images without the fine details needed for cytometric profiling as they depend on the weighted average of available data for inference. We propose a method to overcome this challenge using a generative modeling approach using de-noising diffusion probabilistic models (DDPMs). These models have demonstrated state-of-the-art image generation performance by gradually removing noise.

Our results showed that DDPM outperformed regression-based methods with the best Pearson Correlation Coefficient (PCC) based on the whole image (improved from 0.90 ± 0.04 to 0.95 ± 0.03) and around boundaries (from 0.79 ± 0.07 to 0.88 ± 0.03). In addition, we demonstrate how the focus restoration process improves the image processing steps. For cell contact detection, focus restoration yielded a higher PCC (0.08 ± 0.11 to 0.16 ± 0.19) and lower mean squared error (0.02 ± 0.05 to 0.01 ± 0.02), while for cell detection, the number of cells detected had a lower error (2 ± 4 against 8 ± 14 cells) for the restored images. These results illustrate that diffusion models can be added to the arsenal of tools for focus restoration in high-throughput time-lapse phase contrast microscopy.