Artificial intelligence for stromal cell quality assessment: A systematic review of human mesenchymal stromal cell properties and viability

Introduction

Mesenchymal stromal cells (MSCs) play essential roles in regeneration and repair within the body (Chen et al. 2023; Wang et al. 2025a). MSCs possess immunomodulatory and multipotent differentiation capabilities (Zhu et al. 2025b) and have roles such as in corneal stromal repair and regeneration (El Zarif et al. 2021; Oh et al. 2008; Watson et al. 2010). Mesenchymal stromal cells are attractive candidates for regenerative therapies. Effective clinical translation of stem cell-based therapies requires robust quality control of cultured cells, particularly with respect to viability and maintenance of stemness.

Conventional methods for assessing stem cell cultures rely on manual techniques such as morphological evaluation, biomarker staining, and colony-forming assays (Chen et al. 2023). These approaches are time-consuming, resource-intensive, subject to interobserver variability, and limited in scalability (Chabert et al. 2018). Recent advancements in artificial intelligence (AI), including machine learning and deep learning, offers potential solutions by enabling objective, high-throughput analysis of stem cell characteristics (Mota et al. 2021). There remains limited consolidated evidence on the application and outcomes of AI in regenerative medicine.

This systematic review aimed to (1) systematically summarise artificial intelligence methods used to assess stemness and viability in human MSCs, (2) synthesise the reported performance of these models where appropriate, and (3) critically evaluate their current readiness for stem cell quality assessment and future clinical translation. We synthesized existing studies on AI applications in MSC culture evaluation to identify the types of AI architectures employed, input features analysed, and performance outcomes reported. By mapping the current landscape, this review identifies current evidence gaps and may inform future efforts to standardize AI-based stem cell quality assessment for regenerative medicine. Here, mapping refers to systematically characterising and organising the available evidence according to study characteristics, including MSC source, imaging modality, AI architecture, input features, target outcome, reference standard, and reported performance.

For the purposes of this review, stemness refers to the capacity of MSCs to maintain self-renewal and multipotent differentiation potential while retaining their characteristic biological phenotype. As no single biomarker or functional assay fully defines MSC stemness, the included studies evaluated surrogate indicators including differentiation capacity, senescence status, morphology, proliferation, gene or protein expression, metabolic activity, and other imaging-derived features. Viability refers to the proportion of living, healthy cells capable of maintaining membrane integrity and normal metabolic function. Accordingly, studies assessing viability through morphological analysis, cell segmentation, or other image-based measures of live cell status were considered eligible. These definitions were used to provide a consistent framework for interpreting the diverse AI applications identified in this review. Notably, Silva Couto et al. conducted a side-by-side comparison of adult- and perinatal-derived hMSCs, demonstrating source-dependent biological differences that may affect manufacturing processes and the generalisability of AI models across MSC sources (Silva Couto et al. 2024).

Materials and methods

This systematic review was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines (Page et al. 2021) (Fig. 1). The review protocol was prospectively registered with the International Prospective Register of Systematic Reviews (PROSPERO; Registration No. CRD42025639235)

Results

The number of publications applying AI to MSC stemness and viability has increased in recent years. Of the 18 studies included, 1 was published in 2017, 1 in 2018, 3 in 2021, 2 in 2022, 4 in 2023, 4 in 2024 and 3 in 2025 as detailed in Table 2.

Discussion

Most studies showed low to moderate risk of bias, suggesting the findings are generally reliable but should be interpreted with some caution. A key translational bottleneck is that the mechanisms through which hMSCs exert therapeutic effects remain incompletely understood. This uncertainty makes it difficult to define biologically meaningful potency assays and quality attributes, and therefore to determine whether AI-derived measures of morphology, viability, or stemness reliably predict clinical activity.

The growing number of publications in recent years suggests rising interest in AI-driven stem cell technologies from 1 study in 2017 to more than 3 a year from 2021 onwards (Monson et al. 2024). Most of the studies had been conducted in China or the USA. Supervised models, most commonly CNNs and SVMs were used in almost all studies (17 out of 18). A broad range of MSC sources was represented, predominantly bone-marrow (n = 12) and adipose-derived MSCs (n = 4), with a small number of studies using umbilical-cord or amnion-derived MSCs (n = 1). MSCs have broad potential for regenerative medicine in applications across the body (Chabert et al. 2018), which may explain the strong interest in this area. Most studies assessed morphology, senescence, differentiation capacity, or other surrogate indicators of MSC stemness, while a smaller number evaluated molecular markers.

Our findings suggest that AI could support more consistent MSC quality control by reducing observer-dependent variation and enabling scalable analysis of cell cultures. AI-based image analysis may help identify changes in viability and stemness-associated features earlier and more reproducibly than manual assessment alone. This could streamline culture monitoring and batch-release decision-making; however, AI outputs must be validated against biologically meaningful functional and potency assays before they can be used in clinical manufacturing (Ghabriel et al. 2023). Previous work has also shown that convolutional neural networks can automate cell detection, segmentation, and morphology-based classification from microscopy images, reducing reliance on manual cell counting while maintaining comparable performance (Cadena-Herrera et al. 2015).

The present findings should also be interpreted in the context of the wider MSC translation literature. Although the included studies reported promising proof-of-concept performance, substantial manufacturing barriers remain, including donor-to-donor and tissue-source variability, culture-induced senescence or phenotypic drift, inconsistent media and expansion protocols, scale-up challenges, and the absence of universally accepted potency assays. These factors can alter MSC biological activity and may limit the generalisability of AI models trained on single-site datasets. AI may help address some of these barriers by enabling longitudinal, non-destructive monitoring of culture morphology and quality; however, it should complement rather than replace functional testing. This need for robust and scalable quality assessment is also evident in gene-modified MSC products, where manufacturing processes must preserve cell identity, viability, genetic modification, and therapeutic function during expansion. For example, scalable microcarrier-based manufacturing of gene-modified hMSCs has been investigated to support more reproducible production (Couto et al. 2023), while clinical translation of MSCTRAIL illustrates the need for rigorous product characterisation and potency assessment in ex vivo gene-modified MSC therapies (Graham et al. 2026). Together, these studies support the clinical relevance of developing validated AI tools that can link measurable culture features with functional potency and eventual therapeutic performance.

Heterogeneity was a central finding rather than simply a methodological limitation. Studies differed in MSC tissue source, donor characteristics, culture conditions, imaging modality, AI architecture, reference standards, and the biological outcomes being predicted. Consequently, an accuracy value reported for morphology-based classification cannot be interpreted as directly comparable with a Dice score for image segmentation, a correlation coefficient for continuous prediction, or an F1 score obtained from an imbalanced classification dataset. This variation likely explains the broad range of reported performance and precluded meaningful quantitative pooling. It also highlights the need for future studies to report datasets, class balance, reference standards, validation procedures, and a core set of task-appropriate performance measures in a consistent manner.

Our findings are consistent with the broader literature showing that automated image analysis can support cell detection, segmentation, and morphology-based characterisation while reducing reliance on manual assessment (Cadena-Herrera et al. 2015). Within the included literature, AI was applied across several distinct but complementary tasks, including identification of culture errors from morphology (Imai et al. 2018), cell segmentation and phenotype classification (Mota et al. 2021), prediction of multipotency across donors (Kim et al. 2022b), and identification of senescent cells from bright-field images (He et al. 2024). Collectively, these studies support the potential of AI to detect biologically relevant culture features. However, most models were developed using relatively limited, site-specific datasets and were evaluated against laboratory surrogate outcomes rather than clinical potency or therapeutic efficacy. The present review therefore suggests that AI-based MSC assessment is promising, but remains at a proof-of-concept stage rather than being ready for routine clinical manufacturing.

Limitations of the review include study heterogeneity which precluded meta-analysis, and reliance on English-language publications. Our search strategy did not include the terms ‘umbilical cord’ or ‘Wharton’s jelly’. As terminology for umbilical cord-derived MSCs is not fully standardised, relevant studies may have been missed despite the broader MSC search terms used. While accuracy, sensitivity and F1 were common, some studies used less standard metrics such as LOO-R² or PCC, making comparisons difficult. Reported performance metrics varied substantially across studies owing to differences in AI models, imaging modalities, datasets, and outcome definitions. For example, in one study precision ranged from as low as 21.6% (Mota et al. 2021) which may have been due to imbalanced datasets aimed at detecting rare culture errors, to over 90% in CNN-based stemness predictions (Kim et al. 2022b). These wide ranges, driven by differences in AI models, imaging inputs, and outcome definitions, make direct comparisons across the studies difficult and further underscore the need for standardized evaluation protocols with consistent metrics.

Future research should focus on standardizing outcome reporting of AI models for MSCs (2024). Such research should consider practical and ethical issues related to data accessibility, privacy, governance, and ethical oversight. Many AI models require large, well-labelled datasets (Li et al. 2025), such that most stem cell labs may need to generate and label their own data. Ethically, using patient-derived cells would require strict adherence to consent and privacy. Dedicated studies are also needed for other stem cell types to address the current gap, for example Limbal Epithelial Stem Cells cultures could be assessed with AI given their central role in corneal regeneration. In addition, external validation of AI models for stem cells with independent datasets and the use of shared benchmarks will be critical for translating AI from lab-based models into clinically robust tools for regenerative medicine.

Importantly, most included studies represented early-stage proof-of-concept investigations rather than clinically validated AI systems, and therefore their reported performance should be interpreted cautiously.

Conclusion

This systematic review found that AI shows promising potential for non-destructive assessment of mesenchymal stromal/stem cell culture characteristics, including morphology, confluence, senescence, viability, and stemness-related features. However, the evidence remains heterogeneous and largely limited to early proof-of-concept studies using small, site-specific datasets and laboratory surrogate outcomes. Before AI can be adopted in routine MSC manufacturing or clinical translation, models require standardised outcome definitions and reporting, independent external validation, and evaluation against meaningful measures of cell potency and therapeutic function. AI should therefore be viewed as a potentially valuable adjunct to, rather than a replacement for, established functional and quality-control testing.

Supplementary information

Below is the link to the electronic supplementary material.

Author contributions

W.L. conceived the study, designed the review, conducted the literature search, performed data analysis, interpreted the findings, and wrote the main manuscript. Y.K., S.A., and S.A. assisted with data extraction and risk-of-bias assessment. H.K., D.M., D.B., Y.K., S.A., S.A., and S.B. critically reviewed and edited the manuscript. All authors read and approved the final manuscript.

Funding

Open Access funding enabled and organized by CAUL and its Member Institutions Author Himal Kandel is funded by the Kornhauser Research Associate grant.

Data availability

No datasets were generated or analysed during the current study.

Declarations