Flow cytometry has undergone a dramatic technological expansion over the past three decades. The field progressed from 3-color systems in 1991, to 17-color panels in the early 2000s, to spectral cytometers in the early 2010s, and now to 40-plus color configurations capable of simultaneously measuring dozens of cellular markers on millions of individual cells per sample. Combined with increasing clinical throughput and research volumes, this expansion means that today's flow cytometry labs generate data at a scale that simply cannot be interpreted by traditional manual gating methods. Machine learning (ML) has emerged as the necessary response to this data deluge.
What this review covers: Published in Cancers (2025) by researchers from the University of Utah and ARUP Laboratories' Division of Applied Artificial Intelligence, this review provides a comprehensive introduction to ML techniques as applied to clinical flow cytometry. The authors, who have direct experience deploying ML models in a functioning clinical laboratory, organize their review around three major learning paradigms: supervised learning, unsupervised learning, and weakly supervised or semi-supervised methods. They then address the practical reality of clinical implementation, including regulatory considerations, infrastructure choices, software engineering requirements, and validation strategies.
Why it matters for lymphoma and hematologic malignancies: Flow cytometry is the primary diagnostic tool for immunophenotyping hematologic malignancies. Distinguishing between closely related B-cell neoplasms, detecting minimal residual disease (MRD) in acute leukemia, and identifying rare tumor populations in a background of normal cells are all tasks where computational assistance can meaningfully improve diagnostic accuracy and throughput. The paper covers several published ML systems for detecting conditions including acute myeloid leukemia (AML), chronic lymphocytic leukemia (CLL), Hodgkin lymphoma, and various B-cell neoplasms, while also charting a path toward broader clinical deployment.
One key point the authors emphasize from the outset is that flow cytometry data is inherently digital in format, stored in the Flow Cytometry Standard (FCS) file format first introduced in 1990. This means that, unlike FISH or gel electrophoresis, flow cytometry outputs do not require a separate digitization step before computational analysis can begin. Python packages like FlowKit/FlowIO and the R package flowCore provide direct access to FCS data, after which spectral compensation, logarithmic or arcsinh transformations, and quality control steps are applied before any ML pipeline processes the events.
Supervised machine learning requires a ground truth label for every training example, typically provided by a pathologist's final diagnosis. Once trained, these models predict labels for new, unseen samples. For flow cytometry, the labels most commonly represent disease states (AML present or absent, CLL vs. other B-cell malignancy, etc.) or cell-level identities. The authors provide a detailed taxonomy of supervised algorithms, with practical guidance on when each is appropriate for clinical flow cytometry datasets, which typically range from roughly 100 to 100,000 training examples.
Support Vector Machines (SVMs): SVMs find a decision boundary (hyperplane) that maximizes the margin between class examples. Kernel methods extend this to non-linear boundaries by transforming data to a higher-dimensional space. In a landmark 2015 study from the University of Washington, the authors' colleague David Ng transformed raw cytometry event data into 2D histograms of all pairwise marker combinations and applied an SVM classifier to detect classic Hodgkin lymphoma. The AHEAD group took a different approach, fitting Gaussian Mixture Models (GMMs) to extract specimen-level features before using those features as SVM input for leukemia classification.
Ensemble methods: Random Forests and XGBoost: Random forests build an ensemble of decision trees, each trained on a bootstrapped data subset with random feature subsets, then classify by majority vote. Gradient-boosted trees (specifically XGBoost) construct the ensemble sequentially, with each new tree correcting the residual errors of the previous ones. XGBoost is described as the "de facto standard" for supervised ML on tabular datasets. The review notes that these methods are fast, computationally inexpensive, and challenging to outperform at the dataset sizes typical of clinical flow cytometry. The authors' own group deployed an XGBoost classifier built on Self-Organizing Map (SOM) projections to classify AML, representing what they describe as the first fully deployed ML model in a clinical flow cytometry laboratory. Random forests were also used to classify various B-cell malignancies using UMAP projections of raw flow cytometry data, and to identify Hodgkin lymphoma from a limited dataset.
Neural Networks: Neural networks excel at capturing complex non-linear relationships but require larger training datasets and more hyperparameter tuning compared to ensemble methods. For tabular flow cytometry data, performance is typically comparable to gradient-boosted trees. Simonson et al. transformed raw flow data to 2D histograms and applied convolutional neural networks (CNNs) to detect Hodgkin lymphoma. Zhao et al. used SOMs to aggregate cell data into specimen-level representations fed to a neural network to classify mature B-cell neoplasms at what the authors called "hematologist-level" performance. Hu et al. applied neural networks directly to mass cytometry data without intermediate aggregation, adding a simple pooling operation for specimen-level detection of cytomegalovirus infection. The Mayo Clinic group generated cell-level annotations on approximately 200 patient samples to train a neural network cell classifier for MRD testing in chronic lymphocytic leukemia.
Most supervised ML methods applied to specimen-level flow cytometry tasks require a preliminary unsupervised step to convert raw cell-level event data into a fixed-size feature vector that can serve as model input. This two-stage approach adds complexity and can introduce information loss. Multiple Instance Learning (MIL) sidesteps this by treating the entire specimen as a "bag" of individual cells (instances), with labels assigned only at the bag (specimen) level rather than requiring annotations for each individual cell.
How MIL works: In the attention-based MIL framework described by Ilse et al. (a foundational paper in the field), an initial neural network block transforms each cell into a latent representation, a second attention-based aggregation block learns which cells are most predictive of the specimen label and weights them accordingly, and a final classification block predicts the bag-level outcome. Crucially, the entire pipeline is one end-to-end trainable model, so the aggregation function learns jointly with the feature extraction and classification components. This elegance means that MIL simultaneously learns which cells matter and how to classify based on them, without requiring separate unsupervised featurization.
Clinical application in leukemia: Lewis et al. (2024) applied an attention-based MIL model to flow cytometry data for the detection and molecular characterization of acute leukemias. The model was able to reliably discriminate AML from B- and T-lymphoblastic leukemias and also predict the presence or absence of specific cytogenetic aberrancies and genetic variants using flow cytometry data alone, without requiring molecular testing for those predictions. The authors note that MIL's ability to predict genetic features from phenotypic data represents a particularly exciting potential application.
Future prospects and challenges: The review authors describe MIL as "a very promising approach for future work" in flow cytometry, noting that beyond the Lewis et al. study, they are unaware of other published MIL applications to flow data. The main technical challenge is designing good instance encoders for flow cytometry data. Unlike computer vision MIL models, which can use powerful off-the-shelf pre-trained vision models as instance encoders, flow cytometry lacks analogous pre-trained foundation models for cell-level encoding, requiring researchers to develop these from scratch. The success of MIL in histopathology (where it powers many whole-slide image classifiers) suggests that this architecture transfer could yield significant performance gains once appropriate cell encoders are developed.
Unsupervised learning requires no labeled training data, instead finding structure inherent in the data. In flow cytometry, this is particularly valuable for identifying novel or rare cell populations, characterizing immunophenotypic patterns without prior hypotheses, and reducing the high-dimensional marker space to interpretable representations. The review covers a wide range of unsupervised methods, with careful attention to their practical performance characteristics in clinical flow cytometry settings.
Clustering algorithms: K-means is the simplest approach, iteratively assigning cells to K centroids to minimize within-cluster variance. The flowPeaks R package applies K-means to flow data, while the web application Freecyto uses weighted K-means to enable interactive large-dataset analysis. DBSCAN (Density-Based Spatial Clustering of Applications with Noise) groups densely packed cells without requiring a preset number of clusters, with FlowGrid enhancing scalability by partitioning the feature space into a grid, enabling clustering of millions of cells in seconds rather than minutes. Gaussian Mixture Models (GMMs) assign probabilistic cluster memberships, with the PhenoGMM tool characterizing microbial populations and the Tailor algorithm applying GMM to handle low-resolution panels with heavy-tailed distributions. The AHEAD group used GMM followed by Fisher vectorization and an SVM classifier as a full end-to-end pipeline. The authors note that all clustering approaches suffer from sensitivity to initialization, data drift, and hyperparameter choices.
FlowSOM and Self-Organizing Maps: FlowSOM is a purpose-built tool for high-dimensional flow cytometry data that uses a Self-Organizing Map (SOM) to create a grid of nodes representing the data topology, then connects them via a minimal spanning tree to visualize cluster relationships, with meta-clustering applying hierarchical clustering to refine biologically meaningful groupings. The review authors' own group used an SOM-based model to predict AML and deployed it clinically for triaging positive samples, which they describe as the first published report of a deployed algorithm in a clinical flow cytometry laboratory.
UMAP and t-SNE for visualization: UMAP projects high-dimensional data to lower dimensions while preserving both local and global structure using a nearest-neighbor graph and cross-entropy minimization. It offers near-linear computational complexity O(~n), making it relatively scalable. Van den Akker et al. adapted UMAP to characterize phenotypic patterns associated with NPM1 mutations in AML using clinical flow cytometry data. t-SNE maps high-dimensional distances to probabilities in a lower-dimensional space but suffers from O(n2) computational complexity. The opt-SNE optimization automates perplexity and learning rate selection to handle datasets with millions of cells, and fltSNE uses Fast Fourier Transformations to achieve claimed O(n) complexity. The authors note that UMAP and t-SNE are better suited to visualization and hypothesis generation than to direct diagnostic deployment, due to their non-linear scaling and inference inefficiency.
Generating high-quality labeled training data in clinical flow cytometry is expensive, time-consuming, and dependent on expert pathologist availability. Fully supervised methods require labels for every training example, while fully unsupervised methods generate results that require expert interpretation before any clinical utility can be extracted. Weakly supervised methods occupy the productive middle ground, leveraging partial labeling or inferred labels to train models at scales that would be impractical with full supervision.
Semi-supervised learning: Semi-supervised approaches typically use pseudo-labeling, where highly confident model predictions on unlabeled data are added to the training set between epochs, allowing the model to iteratively expand its effective labeled dataset. SCINA is a published example specifically designed for single-cell and bulk genomic data subtyping using a semi-supervised approach. This framework is directly applicable to flow cytometry where a laboratory may have a large volume of samples but expert-labeled cases for only a subset of disease categories.
Self-supervised learning (SSL): SSL builds representation spaces without explicit labels by exploiting the inherent structure of the data. Teacher-student frameworks like BYOL (Bootstrap Your Own Latent) use augmented views of the same data to train a student network against a slowly updated teacher network, producing generalizable cell-level embeddings without any disease labels. SimCLR (Simple Contrastive Learning of Visual Representations) and DINOv2 are other SSL frameworks from the computer vision domain that the authors suggest could be adapted to single-cell data. Contrastive learning builds metric spaces where similar cells are pulled together and dissimilar cells pushed apart, a potentially powerful paradigm for learning cell identity embeddings from large unlabeled flow cytometry datasets.
Generative models and foundation models: Variational autoencoders (VAEs) encode data into a probabilistic latent space and can serve as anomaly detectors by flagging samples with high reconstruction error. Transformer architectures, originally designed for natural language, are increasingly applied as foundation models pre-trained on large unlabeled datasets, with downstream fine-tuning on smaller labeled datasets. The authors identify foundation models and transfer learning as a major future direction for flow cytometry, analogous to how pre-trained foundation models have transformed computational pathology for whole-slide image analysis. The key challenge is the absence of a large, diverse, public flow cytometry dataset on which to pre-train such a model.
Clinical implementation of ML in flow cytometry involves substantially more than model development. The review dedicates considerable attention to the engineering, regulatory, and operational challenges that separate an academically published algorithm from a functioning clinical tool. This section draws directly on the authors' own experience deploying the first reported clinical ML system in a flow cytometry laboratory at ARUP Laboratories, giving the guidance an unusual degree of practical credibility.
Software and infrastructure requirements: All clinical ML models require supporting software to execute correctly, store results, handle errors, provide logging, manage permissions and authentication, and display results for laboratorians. The authors state that this supporting software "often has a scope and complexity greater than the model itself" and should be handled by professional software engineers, not data scientists. A product manager is needed to coordinate requirements between pathologists, data scientists, and engineers. Ongoing management of model versions across development, certification, and production environments requires dedicated MLOps (Machine Learning Operations) personnel for sophisticated labs running multiple models. The choice between cloud-based and on-premises infrastructure depends on lab size and stage: cloud is preferred for production due to scalability and reliability, while on-premises hardware is more practical for iterative model development.
Analytical and functional validation: The review describes two validation tasks required before clinical deployment. Analytical validation evaluates model sensitivity and specificity on a held-out dataset, with the critical recommendation that this validation set should be temporally closer to current clinical practice than the training data, to capture any data drift that may have occurred. Functional validation involves end-to-end (ETE) testing from FCS file generation on the cytometer through prediction serving to the reporting system. The authors describe ETE testing as "the most robust method" for validating data pipelines, functioning like a "wet dress rehearsal" that uncovers unusual error modes not detectable by unit testing. After validation, models are packaged as containerized modules (e.g., Docker containers) to ensure portability and isolation from environmental changes like package updates.
Regulatory landscape: The regulatory framework for laboratory-developed tests (LDTs) using AI/ML remains unclear. FDA guidance to date has focused on commercial products submitted through the PMA or 510(k) pathway, leaving LDT implementation in a gray area. As of November 2024, the New York State Department of Health Clinical Laboratory Evaluation Program had not issued significant guidance specifically addressing AI/ML tools in flow cytometry. The authors note that data drift, which occurs when real-world clinical data diverge from the distributions present during training, is a primary cause of performance degradation over time and must be actively monitored and addressed post-deployment.
While the primary clinical motivation for ML in flow cytometry is diagnostic accuracy and operational efficiency, the authors dedicate a substantial section to the use of these methods for biological discovery. High-dimensional flow cytometry data collected at clinical scale contains a wealth of information about cell population structure, rare phenotypes, and disease biology that cannot be efficiently extracted by traditional analysis. ML approaches offer a systematic path toward hypothesis generation and the identification of previously unrecognized cell populations with diagnostic or prognostic significance.
Rare cell population discovery: Traditional manual gating is constrained by predefined panel designs and analyst expectations, making it prone to missing subtle phenotypic variants. Clustering algorithms like FlowSOM have been applied to detect rare populations and minimal residual disease in AML. UMAP has been used to identify phenotypic shifts associated with disease progression and treatment response in leukemia patients. A particularly compelling example is the work of Hu et al., who trained a deep learning model to predict latent cytomegalovirus (CMV) infection from mass cytometry data and, through model interpretation, discovered a highly predictive population of CD8+, CD94+, CD27- T lymphocytes in CMV-positive cases, a biological finding with direct immunological implications. Evrard et al. similarly used unsupervised analysis to uncover previously unrecognized heterogeneity in memory T cell populations at baseline and in response to inflammatory processes.
Genetic prediction from phenotypic data: MIL has been applied to predict specific cytogenetic aberrancies and genetic variants from flow cytometry immunophenotypic data alone, without molecular testing. This cross-modal prediction capability suggests that certain genetic alterations leave reproducible phenotypic signatures detectable in flow cytometry, a hypothesis that would require orthogonal molecular validation but could ultimately reduce the need for expensive genetic testing in selected clinical contexts.
Multimodal integration: The authors identify the integration of flow cytometry with complementary modalities as an emerging frontier. Correlating flow cytometry immunophenotypes with next-generation sequencing, immunohistochemistry, and morphology-based diagnostics within transformer-based architectures and foundation models could provide a more comprehensive characterization of pathophysiology than any single modality alone. Clinical datasets at institutions with deployed ML pipelines, which the review notes can reach thousands to tens of thousands of patients (1-2 orders of magnitude larger than most discovery datasets of dozens to hundreds), are particularly well-positioned for this type of multimodal analysis.
Data quality and preprocessing sensitivity: The performance of all ML methods in flow cytometry depends fundamentally on upstream data quality. Choice of distance metric has a dramatic effect on clustering results: Euclidean distance favors absolute marker intensities, while cosine similarity captures relative expression patterns. Log, arcsinh, and logicle transformations alter the distance between points and change cluster boundaries, and there is no gold standard for which transformation is optimal across panels, instruments, or disease contexts. Spectral compensation errors propagate through the entire analysis pipeline, and inter-cytometer variability, lot-to-lot reagent variability, and temporal drift in instrument performance all represent sources of batch effects that can degrade generalizability. The authors' experience is that clustering techniques in particular see performance decreases when inter-cytometer and temporal data homogenization are suboptimal.
Interpretability and the black-box problem: Deep learning models, particularly those processing individual cell events or using attention-based aggregation, do not provide intuitive diagnostic reasoning accessible to pathologists. SHAP (SHapley Additive exPlanations) values and attention mapping offer partial transparency, but these tools show which features or cells the model prioritized without explaining the underlying biological rationale. Discovery-driven applications using deep learning face the additional challenge that identified patterns or populations may lack a straightforward biological interpretation, requiring orthogonal validation through molecular or experimental studies. The authors call for closer collaboration between computational scientists, experimental biologists, hematopathologists, and clinicians to ensure that ML-driven discoveries translate into actionable clinical or scientific insights.
Infrastructure and expertise costs: The review is direct about the cost of clinical ML deployment: it requires teams of data scientists, software engineers, and MLOps personnel, and for many institutions this investment may be prohibitive. The authors note that in their experience, the majority of effort in deploying a clinical ML pipeline is devoted to infrastructure, documentation, and validation, not algorithm development. The dearth of practitioners with both flow cytometry subject matter expertise and ML engineering skills is a significant bottleneck. Most academic publications focus on algorithm development and do not document the full scope of engineering required for clinical deployment, creating an unrealistic picture of implementation difficulty.
Future directions: The review identifies foundation models pre-trained on large-scale flow cytometry data, analogous to those developed for whole-slide image analysis (CTransPath, UNI), as the most impactful unmet need in the field. Federated learning approaches that enable multi-institutional model training without sharing patient data could address the small, siloed dataset problem. The authors express hope for a virtuous cycle: ML investments generating operational efficiencies, which free resources for discovery research, which identify novel biological features that form the basis of next-generation assay designs. Critically, they argue that only institutions willing to invest in clinical ML infrastructure will be positioned to realize these compounding benefits.