Download

Download Poster (3.4 MB)

Student Class

2029

Student Affiliation

Thayer School First Year Research in Engineering

First Advisor

Irene Georgakoudi

First Advisor Department

Department of Engineering Sciences—Thayer School of Engineering

Second Advisor

Matthew Lindley

Second Advisor Department

Geisel School of Medicine

Abstract

Accurate single-cell segmentation of cytoplasm from nucleus in autofluorescence images is critical for quantifying cellular metabolism using the optical redox ratio (FAD / FAD+NAD(P)H). However, automated cytoplasmic segmentation of T cell two-photon excitation fluorescence images remains challenging due to low signal-to-noise ratio, small cell size, and variability in cell brightness across imaging conditions. Here, we present a pipeline combining a custom-trained Cellpose-SAM model with an adapted cytoplasmic post-processing algorithm (CPPA) for automated single-cell metabolic analysis of T cell NAD(P)H autofluorescence images. Raw 2048×2048 pixel images were spatially binned 4×4 in MATLAB to improve signal-to-noise ratio and reduce processing time. A custom Cellpose-SAM model was trained on five manually corrected T cell regions of interest spanning multiple experimental conditions and validated on three unseen images, detecting between 110 and 262 cells per image. CPPA was implemented in Python and applied to Cellpose masks to separate cytoplasm from nucleus using per-cell intensity normalization, percentile-based thresholding, distance transform edge correction, and connected component nucleus cleanup. Six CPPA configurations were evaluated by comparing per-cell redox ratios across three test images. Per-cell normalization outperformed image-level thresholding, and a 60th percentile threshold with dim-cell correction produced the most visually accurate cytoplasm and nucleus masks. This pipeline enables automated, single-cell metabolic analysis of T cell autofluorescence images and provides a foundation for characterizing metabolic heterogeneity across immune cell populations.

Publication Date

2026

Disciplines

Bioimaging and Biomedical Optics

Training a Cellpose Model for T-Cell Cytoplasm and Nucleus Segmentation

Share

COinS