Implementation of artificial intelligence to automate physical disector in a fractionator design for quantification of stem cell-derived neurons

Accurate quantification of stem cell-derived dopaminergic neurons is essential for advancing cell therapy strategies in Parkinson's disease (PD). Traditional manual stereological methods, while robust, are time-consuming and subject to interobserver variability, limiting their scalability for preclinical and translational studies. This study presents the development and validation of an artificial intelligence (AI)-assisted physical fractionator workflow for unbiased and efficient quantification of human embryonic stem cell (hESC)-derived ventral midbrain dopaminergic (vmDA) neurons in a Parkinsonian rat model. The workflow integrates convolutional neural networks (U-net and DeepLabv3+) within the Visiopharm platform to automate tissue alignment, graft region identification, and cell segmentation/classification based on triple immunofluorescent labelling (TH, FOXA2, HNA). Human-in-the-loop review ensures quality control and allows for flexible adjustment of immunostaining thresholds. Performance was evaluated using paired datasets: brains from Study A (training, validation, and test set) and Study B (out-of-sample test set from a separate experiment). The AI-assisted workflow demonstrated segmentation and counting accuracy comparable to human experts, with high precision, sensitivity, and high F1 scores for both segmentation and quantification. Results showed no significant differences between AI-assisted and manual quantification of total human cells and vmDA neurons across studies, and the workflow substantially reduced hands-on analysis time from 8 h to 1 h per graft. The investment required for AI model development, annotation, and optimization in the initial phase took several months and is regarded as a one-time infrastructure investment. Additionally, the workflow's design enables adaptability for other cell types by integrating relevant markers. These findings highlight the potential of AI-assisted stereological workflows to accelerate and standardize cell quantification in preclinical research, with potential relevance for translational research settings in cell therapy development.