(C-100) Cellpose Reliably Segments Single Cells from Autofluorescence Images

Introduction:: Automatic cell segmentation has greatly improved in recent years due to advances in machine learning and computer vision. These tools offer fast and accurate segmentation while eliminating interobserver variability, though often struggle with label-free, autofluorescence images. Cellpose, a U-Net based algorithm, is computationally efficient, fast, and capable of handling multiple imaging modalities¹. This study aims to characterize the strengths and limitations of Cellpose segmentation models compared to manual segmentation of NAD(P)H autofluorescence images of cells in culture using Dice similarity scores and two-photon (2P) autofluorescence intensity and lifetime imaging variables.


Materials and Methods:: Panc1 cells were treated with either 2DG (10mM; 120 min.) or NaCN (4mM; 15 min.), inhibitors of glycolysis and the electron transport chain, respectively. Cells were promptly imaged with a 40x/1.15 NA objective on a custom inverted 2P-microscope with sequential acquisition of NAD(P)H [750nm ex. (440/80nm em.)] and FAD [890nm ex. (550/100nm em.)]. Cellpose models for single-cell segmentation of nuclei and whole cells were developed by the Center for Open Bioimage Analysis using preexisting manually segmented masks. Dice scores were calculated on the whole image level for a comprehensive comparison of two sets of masks where a Dice score of 0.76 is the two-observer maxima¹. Autofluorescence features including NAD(P)H and FAD mean fluorescence lifetimes (τ_m)_, and optical redox ratio (ORR) were averaged for each nucleus.


Results, Conclusions, and Discussions:: Cellpose and human segmentation provide similar NAD(P)H τ_m values across 2D cultures of Panc1 cells (Fig. 1A) with similar treatment-induced changes in NAD(P)H τ_m for Cellpose and human segmentation (Fig. 1B). Reproducibility of manual human segmentation and Cellpose segmentation was tested by performing segmentation, rotating the same image by 90°, and performing segmentation again. Dice coefficients show that manual human segmentation (M1-M2) is less reproducible than Cellpose segmentation (C1-C2), with the manual to Cellpose comparison (M1-C1) similar to the manual-manual comparison (Fig. 1C). Cellpose can also segment more cells in less time than manual segmentation.
Cellpose generates masks comparable to manual masks for label-free autofluorescence images. We have also confirmed generalizability by applying Cellpose to segment single cells within autofluorescence images of 3D patient-derived cancer organoids. Cellpose performance depends on the quality of the training masks, though further refinement can be achieved by incorporating human feedback, creating more desirable masks over time. Overall, Cellpose is fast, reproducible and shows promise as a reliable method for segmentation in label-free microscopy.


Acknowledgements (Optional): :

References (Optional): : 1. Stinger, S. Nat. methods. 2020.