ABS 044: Leveraging Deep Learning for Enhanced 3D Confocal Imaging of Circulating Hybrid Cells

Suin Jung ¹ ², Sandhya Govindarajan ², William Greer ², Abigail Moore ³, Melissa Wong ³ ⁴, Summer Gibbs ² ⁴, Young Hwan Chang ² ⁴

¹ Brown University
² Department of Biomedical Engineering, Oregon Health & Science University
³ Department of Cell, Developmental and Cancer Biology, Oregon Health & Science University
⁴ Knight Cancer Institute, Oregon Health & Science University

The Van Wickle Journal (2026) Volume 2, ABS044

Introduction: Circulating Hybrid Cells (CHCs) are tumor-derived cells in the bloodstream that evade the body's immune response by co-expressing tumor and immune markers, making them promising biomarkers for cancer progression and metastasis. Most current CHC studies rely on 2D imaging, which cannot fully capture cellular morphology or spatial marker distribution, limiting phenotypic resolution. To overcome these limitations, we utilize 3D confocal microscopy to capture full Z-stacks, providing richer spatial subcellular information; however, this approach requires intensive data acquisition. We address this challenge with deep learning, utilizing frame interpolation algorithms to generate intermediate Z-slices, thereby reducing the number of physical slices required. This approach aims to lower costs, streamline 3D imaging, and expand its accessibility for robust clinical diagnostics and scalable biological imaging.

Methods: Four murine colorectal cancer-immune hybrid cell lines (H3, H4, H7, H11), derived from clones of the MC38 parent line generated by co-culturing RFP-tagged tumor-like and GFP-tagged macrophage-like cells, were stained with DAPI and antibodies targeting GFP, RFP, CD45, and EGFR. Confocal z-stacks of 16-22 slices were acquired across cell depth at 0.6 μm step size. To increase z-resolution, we adapted Real-time Intermediate Flow Estimation (RIFE), a deep learning model leveraging a U-Net-like architecture for optical flow estimation, to interpolate intermediate z-slices from pairs of adjacent input slices. Model performance was evaluated using Structural Similarity Index Measure (SSIM), comparing predicted intermediate slices against held-out ground truth slices, with reconstruction loss used to iteratively refine interpolation quality. Single-cell segmentation was performed using MESMER, enabling mean intensity correlation analysis at single-cell resolution across fluorescent channels.

Results: RIFE enabled up to 75-80% reduction in z-slice acquisition while maintaining SSIM scores >0.98 across all five fluorescent markers (DAPI, GFP, RFP, CD45, EGFR). Difference maps of interpolated versus ground truth slices confirmed minimal perceptible difference between the slices across reduction levels. Single-cell mean intensity correlations further demonstrated that biologically relevant signal is preserved at single-cell resolution, validating that z-stack frame reduction does not compromise downstream quantitative analysis.

Discussion: Our research demonstrates that deep learning-based z-slice interpolation using RIFE can significantly accelerate confocal imaging workflows without compromising the biological integrity of cellular structures. By preserving high structural similarity and single-cell intensity correlations across all markers, RIFE offers a practical path to significantly reducing imaging time, storage, and cost in high-throughput settings. Future work will extend validation to additional hybrid cell line clones and patient-derived samples to assess clinical translational potential and model generalizability beyond murine systems.