Adapting Cellpose for Segmentation and Quantification of Cultivated Limbal Stem Cells

Abstract

Purpose : Cultivated limbal epithelial cells are currently being explored as a potential treatment for limbal stem cell deficiency (LSCD). Part of the quality control process requires manual quantification of cell features such as cell density, which correlates with progenitor cell content. Identifying cultured limbal epithelial cells in culture media and quantifying relevant biomarkers requires a trained reader to manually annotate microscopy images in a time-consuming process. An automated approach for identifying morphological features of cultured limbal epithelial cells will streamline the quality control process and reduce interrater variability.

Methods : Our proposed machine learning algorithm builds upon the pretrained Cellpose V2 generalist cell segmentation model. Using a dataset of over 800 microscopy images with cell locations annotated with ROI points, we finetuned the model to improve its segmentation performance on cultured limbal epithelial cells. Cellpose typically requires the segmentation of whole cell boundaries as training data for fine-tuning. To use our pre-existing data, we generate a training mask by passing the original image through a pretrained Cellpose model to create an initial set of masks that are matched to corresponding point ROIs. Unmatched masks are removed and artificial masks of a fixed radius are assigned to unmatched point ROIs. This training data is used to iteratively refine the model until it converges.

Results : Initial evaluation on the test set (n=266) using this new approach has shown that the mean absolute errors (MAE) of the cell counts from the base Cellpose model (19.49) and the fine-tuned Cellpose model (19.96) are lower than the previous computer vision method (34.54). When evaluated as a classification task for cell grading, the base Cellpose and fine-tuned Cellpose models achieve higher accuracies of 86.46% and 93.98%, respectively, compared to 60.75% of our previous computer vision method. While the base Cellpose model had slightly better performance in MAE, the fine-tuned Cellpose model has a higher classification accuracy.

Conclusions : We have adapted an existing generalist cell segmentation algorithm to perform cell segmentation on an imprecisely labeled dataset. Additional data are needed to further refine the model and validate prospectively.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.