Toward integrated multimodal artificial intelligence in bone marrow examination: current state and open challenges

Abstract

Bone marrow (BM) examinations are intrinsically multimodal: hematopathologists integrate peripheral blood, aspirate cytomorphology, trephine histology, flow cytometry, cytogenetics, and molecular genetics data into a single interpretation. Because WHO-HAEM5 and ICC 2022 define entities through integrated morphological, immunophenotypic, cytogenetic, and molecular criteria, multimodal diagnostic reasoning embedded in a procedural workflow, not isolated image classification, represents the appropriate target for the use of artificial intelligence (AI). We review the existing research on integrated multimodal AI systems for BM examination by classifying these systems into three categories: mature single-modality tools; systems that perform cross-modal prediction, in which one modality is estimated from another; and genuine multimodal fusion systems, in which several modalities are jointly ingested by a single model. Evidence is stronger at the single-modality level, where deep learning approaches expert performance in aspirate cytology, trephine histology, and flow cytometry assessments, and the first Food and Drug Administration (FDA) De Novo–authorized all-digital bone marrow aspirate (BMA) application has defined a new device category. On the other hand, while cross-modal prediction is rapidly expanding, genuine multimodal fused systems that jointly ingest matched morphology, flow, molecular, and clinical data to yield unified guideline-aligned recommendations remain unrealized. Direct head-to-head evidence is sparse and endpoint-dependent; nevertheless, the lack of demonstrated superiority over the strongest single-modality comparator does not automatically imply the absence of benefits. The dominant obstacles are fusion of heterogeneous modalities with missing data, absence of large paired multi-institutional datasets and harmonized nomenclature, class imbalances, batch-effect-driven generalization failures, immature explainability, uncertainty quantification, and regulatory pathways. Nevertheless, integrated multimodal AI systems for BM examinations have entered a promising early translational phase, and the field should aim to pursue integration using paired datasets, explicit fusion architectures, calibrated abstention-aware models, and staged prospective validation.

Introduction

Bone marrow (BM) examinations, including analyses of bone marrow aspirate (BMA) smears and trephine core biopsy (BMB), are the gold standard for the diagnosis, classification, staging, and therapeutic monitoring of hematological malignancies and benign marrow-failure syndromes. Despite the availability of mature ancillary techniques such as flow cytometry, cytogenetics, and next-generation sequencing, morphological evaluation is indispensable, providing spatial, architectural, and cytological contexts that cannot be obtained from liquid or homogenized samples. Historically, however, such interpretations have been laborious, subjective, and expertise-dependent, with well-documented inter- and intra-observer variability when grading subtle dysplasia, quantifying rare cells, and estimating cellularity.

Two developments have reshaped this landscape. First, the digitization of pathology through whole-slide imaging (WSI) has enabled the systematic application of AI in routine hematological assessments, even extending it to generative large language models in research and reporting workflows [1]. Second, and more consequentially, contemporary disease-classification frameworks have explicitly characterized the multimodal nature of marrow diagnosis: The fifth edition of the World Health Organization (WHO) classification of hematolymphoid tumors (WHO-HAEM5) and the International Consensus Classification (ICC 2022) have both moved toward genetically defined entities, in which morphology must be interpreted together with data pertaining to the immunophenotype, cytogenetics, and molecular lesions [2–4]. The more the diagnostic reference standard is jointly constituted by these modalities, the stronger the case for integrated multimodal AI and the more clearly the current unimodal successes prove to be insufficient.

A recurring source of confusion is the imprecise use of “multimodal”: many highly cited “multimodal” systems are in fact unimodal image models predicting multimodal clinical truth labels—for example, inferring an NPM1 mutation from a digitized smear. This review categorizes systems using an explicit tripartite framing: (a) mature single-modality tools; (b) systems performing cross-modal prediction, in which one modality estimates another; and (c) genuinely fused multimodal systems, in which multiple modalities are simultaneously ingested by a single model, with category (c) remaining largely unrealized in BM examination.

This review delineates the state of the art across all three categories of systems, the milestones and bottlenecks that define their present capabilities, and a research roadmap anchored throughout the actual clinical procedure that determines feasible architectures, benchmarks, and evidentiary standards (Fig. 1).

Fig. 1.

Integrated multimodal artificial intelligence for bone marrow examinations: Current status and translational roadmap. a Bone marrow examinations are intrinsically multimodal. In routine clinical practice, peripheral blood findings, aspirate-smear cytomorphology, trephine biopsy histology and immunohistochemistry, flow cytometry, cytogenetics/FISH, and molecular genetics data and clinical metadata are integrated to generate a coherent diagnosis, classification, and risk assessment. b The existing bone marrow artificial intelligence systems can be organized along a maturity spectrum. The evidence is the strongest for single-modality tools addressing cytological examinations, trephine whole-slide image analysis, flow cytometry classifiers, and cytogenetic or molecular-prognostic models. Cross-modal prediction, such as morphology-based prediction of molecular or cytogenetic labels, is a rapidly expanding but conceptually distinct category because the predicted modality is not measured and jointly ingested at inference. Genuine multimodal fusion requires simultaneous incorporation of multiple measured modalities and remains the central unmet goal for the use of artificial intelligence in bone marrow examinations. c A realistic target system is a case-level multimodal AI orchestrator built around specialist encoders for cytology, histology/WSI, flow cytometry, cytogenetics, molecular genetics, and clinical/report data. Such a system should support intermediate or late fusion, handle missing and asynchronous modalities, and provide WHO-HAEM5/ICC-aligned integrated diagnosis with risk estimation, ancillary test prioritization, evidence packets, and calibrated uncertainty or abstention for human-confirmed sign-out. d Translation of such a system to clinical practice will require paired multi-institutional datasets, control of batch effects and drift, missing-modality-aware modeling, harmonized labels and rare-class learning, explainability and calibration, and staged prospective validation. AI, artificial intelligence; BM, bone marrow; CBC, complete blood count; FISH, fluorescence in situ hybridization; ICC, International Consensus Classification; IHC, immunohistochemistry; NGS, next-generation sequencing; PCR, polymerase chain reaction; WHO-HAEM5, fifth edition of the World Health Organization Classification of Hematolymphoid Tumors; WSI, whole-slide imaging

Review search strategy and evidence classification

This narrative review was based on searches of PubMed/MEDLINE and Europe PMC from database inception, which were supplemented by searches using Google Scholar, reference-list screening, and targeted searches of publisher websites and other relevant web sources; the final search was conducted on July 8, 2026. The search terms combined BM examination terms (“bone marrow,” “bone marrow aspirate,” and “bone marrow biopsy”) with AI or machine-learning terms and terms for component modalities, multimodal integration, and cross-modal prediction. Peer-reviewed original studies directly relevant to AI in BM examination or in its component modalities were included. Reviews and guidelines were used for context, while non-hematologic studies, studies in which BM examinations were incidental, and superseded preprints were excluded from core claims.

Evidence classes were assigned on the basis of the measured inputs available to the model at inference: single-modality systems used one measured modality; cross-modal systems predicted a label derived from another modality without ingesting it; partial-fusion systems jointly used two or more measured modalities; and systems with comprehensive integration combined morphology, raw flow cytometry, cytogenetic or molecular data and structured clinical information to generate a unified WHO-HAEM5/ICC-aligned output. Multimodal pre-training followed by single-modality inference and off-marrow examples were treated as the methodological context. For consistency of presentation across studies, the reported performance estimates were rounded only for display: percentages to one decimal place and area under the receiver operating characteristic curve (AUROC), concordance, correlation, and related coefficients to no more than two decimal places; fixed clinical thresholds retained their conventional notation.

Bone marrow examination as a multimodal diagnostic workflow

Any credible target for the use of AI in BM examinations must account for how the examinations are performed and interpreted. The International Council for Standardization in Hematology (ICSH) guidance emphasizes that the final interpretation should integrate peripheral blood, aspirate smear, particle clot, trephine biopsy, and imprint morphology data with the results of cytochemistry, immunophenotyping, cytogenetics, and molecular genetics assessments, where appropriate [5], defining the genuine task of multimodal diagnostic reasoning embedded in a procedural sequence, not the isolated classification of a single image (Fig. 1a).

Aspiration and cytology

The aspirate is the core sample for cytomorphological analysis. The ICSH recommends a small first draw (~0.5 mL, because larger volumes dilute the marrow with blood) and, when precise enumeration is required, at least 500 cells across two smears (or 300 otherwise), increased near a critical threshold, such as 5% blasts [5]. These details are not incidental to AI: label quality, hemodilution, blast percentage, and the myeloid-to-erythroid (M:E) ratio all depend on them, and evaluation of the 300–500-cell differential—the labor-intensive, variability-prone task targeted by early marrow AI systems [6]—is the least reliable in the near-threshold zone, where an incorrect determination changes a WHO/ICC blast-based category. The next-generation-flow bone marrow quality index (BMQI) flags hemodilution and distinguishes first- from second-pull aspirates, supplying the per-sample quality label that AI training sets rarely record [7]. “Rescue” specimens—particle-clot preparations and dry-tap imprints—are diagnostically important but underrepresented in AI datasets, biasing models toward technically adequate aspirates rather than difficult cases in which assistance is most valuable.

Biopsy, histology, and immunohistochemistry

The trephine contributes the architectural view that smears cannot provide, supporting assessments of cellularity, fibrosis, focal infiltrates, granulomas, metastatic disease, lymphoid aggregates, and megakaryocyte topography, which are critical features in myeloproliferative neoplasms (MPNs), myelodysplastic neoplasms (MDS), plasma-cell neoplasms, and lymphoma staging. Particle-clot sections provide complementary information while avoiding decalcification-related nucleic acid and protein damage, with implications for pipelines linking morphology to molecular readouts [5]. Because trephine interpretation is often inseparable from targeted phenotyping, the ICSH also standardizes BM immunohistochemistry [8]: “histology” should not be equated with hematoxylin-and-eosin staining alone, since auxiliary staining procedures and immunophenotypic maps are potentially informative modalities in their own right.

Flow cytometry and molecular tests

The ICSH recommends collecting aspirate material for supplementary studies, including flow cytometry, cytogenetics, fluorescence in situ hybridization (FISH), and molecular genetics [5]. In acute leukemia, standardized panels such as the EuroFlow Acute Leukemia Orientation Tube (ALOT) orient the lineage and guide the work-up [9]. Because flow cytometry is faster and more standardized than morphological assessments for several tasks, it has become a parallel AI development stream. Most importantly, because the WHO-HAEM5 and ICC 2022 genetically define many entities [2, 3], the clinical standard is already integrated, and the question is whether AI can match it.

Foundations: unimodal AI across individual modalities

Before examining fusion, we established the progress within each isolated modality—the unimodal encoders from which integrated systems must be assembled (Fig. 1b). In this regard, the two morphological specimens pose distinct challenges: the aspirate smear resolves individual intact cells, whereas the trephine preserves the spatial architecture of the marrow.

Digital cytomorphology of aspirate smears

Convolutional neural networks (CNNs) dominate BMA cytology by learning features from raw pixels. Early studies established human-level blast recognition in peripheral blood smears [10], and systems such as BMSNet demonstrated hematologist-level smear assistance [11]. The field-defining resource is the Matek dataset (171,374 expert-annotated single-cell images covering 21 cytomorphological classes from 945 patients) [12]. Most commercial cytology systems follow a pre-classification-plus-verification paradigm, exemplified by CellaVision's neural network peripheral blood differential (U.S. FDA Class II) and its BMA application, which pre-classifies aspirate cells on a dedicated analyzer for expert confirmation [13, 14]; the marrow application is available in Europe, though its published validation remains limited. The reported accuracy of such systems is high yet uneven: Morphogo reached 99.0% overall accuracy but only 81.0% average sensitivity, with only moderate agreement for the immature and blast-like classes that drive diagnosis [15, 16]. Clinically decisive failures—misclassifying blasts as lymphocytes and thereby underestimating blast counts—have been documented even for mature peripheral blood preclassifiers [13]. The AIFORIA nine-class classifier likewise paired near-perfect internal precision (≈99.9%) with a ~ 13.0% external blast false-positive rate while omitting mast cells, megakaryocytes, and smudge cells [17]. Such high overall accuracy obscures class-dependent sensitivity and must be read alongside specimen quality and operator review limitations. Deep learning has also been used to classify dysplastic cells across erythroid, granulocytic, and megakaryocytic lineages [18].

A pivotal advancement was the shift from cropped single cells to an end-to-end digital smear workflow. Tayebi et al. built a pipeline operating directly on aspirate WSI (region-detection accuracy: 0.97; mean average precision [mAP] for cells: 0.75), wherein the "Histogram of Cell Types" (HCT) reframed the marrow AI output as a human-readable specimen summary aligned with the clinical differential count [6]. Related frameworks classified 16 cell types at high magnification [19], and slide-level representation learning allowed whole-slide retrieval and label prediction [20].

Clinical translation reached a regulatory milestone with the Scopio Labs X100 with Full-Field BMA, the first AI-assisted digital BMA device to receive U.S. FDA De Novo authorization (DEN230034; decision date: March 22, 2024), establishing a new Class II category (automated cell-locating device for BMA) [21]. Providing a full-field digital review at ~100 × objective-equivalent resolution, this device assists with smear-quality assessment and blast, plasma-cell, and M:E estimation. In multicenter validation, a high overall test–reference agreement was tempered by a fall in the blast-count specificity from acute leukemia to MDS/MPN samples [22]. Critically, the clearance mandates that AI remain assistive; the operator must verify the results, and confirmation of acute leukemia still requires human synthesis.

Trephine biopsy and whole-slide histology

The BMB poses distinct challenges, including densely packed overlapping nuclei and the interplay of hematopoietic tissue, adipocytes, and trabecular bone. Therefore, researchers have to trade annotation costs against precision across WSI supervision strategies, from labor-intensive fully supervised pixel-wise annotation through semi-supervised dot-annotation and weakly supervised multiple-instance learning that scales to slide-level labels, to unsupervised clustering that reveals novel sub-phenotypes at the cost of opaque criteria. The landmark tool MarrowQuant—validated in murine aging and aplasia models [23] and then validated in humans as MarrowQuant 2.0, an open-source QuPath workflow [24]—segments hematopoietic, adipocytic, bony, and interstitial/microvascular compartments and derives cellularity mathematically. Because the naive hematopoietic-over-(hematopoietic-plus-adipose) ratio fails with stromal expansion (fibrosis, gelatinous transformation), it substitutes a bone-referenced formula [24]. The results obtained with this tool tracked hematopathologists’ estimations closely (intraclass correlation coefficient ≈ 0.98), addressing the well-recognized inter-observer variability of cellularity estimation, although its authors cautioned against using the tool in conditions where megakaryocyte hyperplasia is expected, including MPNs [23, 24]. A complementary deep-learning tool that also segments megakaryocyte, erythroid, and myeloid lineages showed only moderate agreement with expert cellularity assessments (intraclass correlation coefficient, 0.78) and systematically underestimated cellularity, indicating that lineage-resolved quantification remains more challenging than binary cellularity [25].

Deep learning in flow cytometry

Flow cytometry is already a structured, high-dimensional modality. Because acquisition produces event-level measurements for individual cells or particles, the data suit set-based and attention methods, and machine learning classifies acute leukemias and distinguishes them from non-neoplastic cytopenias [26]. Pipelines pairing self-organizing maps (FlowSOM) with downstream classifiers have reached 90.0%–97.0% sensitivity and specificity for MDS [27], with an external-validation AUROC of approximately 0.94 [28]. Zhong et al. reported an acute leukemia concordance of 0.98 (κ = 0.96) while reducing the analysis time from 15.6 min to ~84 s [29]. Attention-based multiple-instance learning is especially powerful. Lewis et al. applied it to 1,820 samples (736 acute leukemias), reporting an AUROC of 0.96 for diagnosis and 0.97 for distinguishing acute myeloid leukemia (AML) from B- and T-lymphoblastic leukemia (ALL), and predicting nine cytogenetic aberrancies and 32 variants from flow alone, although the performance was strong only for select lesions (PML::RARA 0.93) [30]. Cheng et al. reported an accuracy of 94.6% for AML and 98.2% for B-ALL (sensitivities of 90.0% and 90.9%, respectively) using EuroFlow ALOT data in 241 patients [31]. Imaging flow cytometry adds a morphological context, discriminating leukemic from normal hematopoietic stem cells with 87.0% accuracy (93.4% with a DNA dye). These findings represent proof of concept for ultrasensitive minimal residual disease (MRD) assessment, although they were obtained in only five AML patients [32]. Some programs have reached deployment: Zuromski et al. validated an AML detection model in a clinical flow laboratory with real-time inference and monitoring [33], and the DUALFLOW study (NCT07269067) will compare unsupervised FlowSOM-based gating with manual gating for AML-MRD.

Cytogenetics, karyotyping, and prognostic modeling

AI has also matured in cytogenetics—metaphase identification, chromosome segmentation, enumeration, and automated karyotyping (e.g., MetaSystems Metafer, KaryoXpert [34], DeepACEv2 [35])—but the systems used for these analyses were validated on non-neoplastic metaphases (peripheral blood and amniotic fluid material for KaryoXpert [34]; routine karyotyping for DeepACEv2 [35]) rather than neoplastic marrow, so their performance on the complex, hypodiploid, or marker-laden karyotypes that dominate myeloid malignancy remains largely untested. On the molecular-prognostic side, the Molecular International Prognostic Scoring System for MDS (IPSS-M) profiled 152 genes across 2,957 subjects, scored 31-gene somatic mutations, re-stratified 46.0% of the patients, and improved the concordance index by approximately 0.05 over the Revised IPSS (IPSS-R) [36]. More recently, machine-learning scores such as the Artificial Intelligence Prognostic Scoring System for Myelodysplastic Syndromes (AIPSS-MDS) have outperformed IPSS-R and matched IPSS-M in external validation [37]. Although not multimodal in the sense of fusion, these systems show that AI has already contributed decisively to the molecular half of the integrated diagnosis.

Hematology foundation models

A decisive shift from ImageNet-pretrained CNNs to domain-specific self-supervised foundation models is currently underway. DinoBloom, the first hematology single-cell foundation model, is a DINOv2 vision transformer trained on 13 public blood and BM smear datasets (> 380,000 images) that provides generalizable feature extractors [38], and a red-cell counterpart (RedDino) has been developed [39]. In adjacent studies on computational pathology, whole-slide models such as TITAN have been shown to align morphology with pathology-report language, whereas mSTAR additionally integrates the gene-expression context, although both are pretrained on pan-cancer repositories that essentially do not represent BM [40, 41]. Cross-attention pathology–genomics architectures allow genomic features to query histological regions [42]. Although not yet marrow-specific, these WSI-scale models are likely to become the default backbone for marrow multimodal systems, especially where locally labeled data are scarce and tasks are rare or fine-grained.

From cross-modal prediction to genuine multimodal fusion

The tools surveyed thus far each operate within a single modality, whereas the diagnoses they inform are assembled from many modalities. A major limitation of traditional workflows is siloed data; pathology, flow cytometry, and genomics data are evaluated independently, forcing physicians to manually synthesize disparate and sometimes contradictory reports. Multimodal AI promises to bridge this gap, but realizing this promise will first require distinguishing what current systems can do.

The critical conceptual distinction

An important clarification is that the systems many studies described as “multimodal” are unimodal image models that predict multimodal clinical truth labels. Predicting a genetic lesion from morphology data (cross-modal prediction) is not the same problem as jointly fusing morphology with measured sequencing output (multimodal fusion). The former regresses one modality onto a label derived from another, while the latter ingests multiple modalities during the inference stage. This distinction determines which claims are defensible and what infrastructure is required; therefore, image-only biomarker prediction should not be labeled “fusion” unless additional modalities are genuinely consumed.

Cross-modal prediction and virtual molecular profiling

Cross-modal prediction, also known as “virtual molecular profiling” or histogenomics, is the fastest-growing category, yet it remains adjunctive. Deep networks detect subtle, sub-visual morphological signatures on standard smears or sections that correlate with underlying mutations, bridging phenotypes with genotypes without sequencing. Eckardt et al. detected AML from BM smears (AUROC: 0.97; 1,251 AML and 236 healthy donor samples) and predicted NPM1 status with an AUROC of 0.92 from image data alone [43]. Kockwelp et al. extended this to an automated pipeline operating on more than two million single-cell images from 408 AML patients, predicting therapy-relevant groups with uneven accuracy: CBFB::MYH11 was well predicted (median AUROC: ~ 0.91), whereas the accuracies for predicting NPM1 status, Medical Research Council (MRC) cytogenetics, and European LeukemiaNet (ELN) 2017 favorable risk were only modest (AUROC: ~ 0.70, ~ 0.68, and ~0.64, respectively), confirming that morphology carries a limited genetic signal for most lesions [44]. Brück et al. linked deep BM histopathology to the clinical and molecular determinants of MDS [45]. These results were promising—an inexpensive morphological screen could shorten the turnaround time—but represented cross-modal inference, not fusion: such systems function as gatekeepers, triggering confirmatory flow cytometry, FISH, or rapid RNA sequencing when digital cytology identifies suspicious patterns.

Algorithmic fusion mechanisms

While cross-modal prediction estimates one modality from another, genuine fusion must simultaneously integrate several modalities by reconciling spatial image data, sequential genomic data, and high-dimensional flow cytometry metrics. Early attempts at “late fusion” simply averaged unimodal predictions; in contrast, the modern state of the art favors multimodal transformers and cross-attention, in which data streams interact in latent space before a final prediction. Three canonical strategies span the design space: early fusion projects all modalities into one joint predictor, which is powerful when modalities are complete and co-registered and brittle when any modality is missing; intermediate fusion encodes each modality separately and then integrates it through cross-attention or gating; and late fusion trains independent models and combines outputs through a meta-learner or rule engine, accommodating asynchronous, missing, or laboratory-specific inputs. Validated templates show that genuine fusion can outperform the best single modality: Deep Orthogonal Fusion outperformed the top unimodal model for glioma survival (concordance index: 0.79 vs. 0.72) in only 176 patients [46]; the Multimodal Co-Attention Transformer gained 3.0%–6.9% concordance across five tumor types [42]; and mSTAR improved on vision-only encoders across 97 tasks [41], but almost entirely outside hematology, on solid tumors with co-registered data. The results obtained with the small Deep Orthogonal Fusion cohort indicate that a fusion advantage does not require enormous datasets. For present-day marrow practice, intermediate or late fusion is often more feasible than early fusion because modalities enter asynchronously (morphology and flow data in one to two days, cytogenetics and sequencing data in one to three weeks), may be missing (dry taps, unavailable genetics), and come from different laboratories; this is one reason the literature contains more “composed” than “fully joint” systems.

Proto-integrated marrow systems

Genuinely fused BM models, including those addressing MPNs and plasma-cell neoplasms, are rare. Head-to-head findings are sparse and endpoint-dependent. Yu et al. reported the same internal macro-AUROC for image-only and comprehensive models (0.96) but a numerically higher average precision with fusion (0.75 → 0.81); only the image-only model was externally validated [47]. Wang et al. found significantly higher AUROCs with fusion for overall MPN-versus-non-MPN and essential thrombocythemia classification, but the AUROC was not consistently higher across other MPN subtypes [48]. ImmunoCast-MM fused cytomorphology and fluorodeoxyglucose F 18 (¹⁸F-FDG) positron emission tomography (PET)/computed tomography (CT) data, with clinical covariates added for outcome prediction; flow cytometry served as an alignment/validation reference rather than an inference input in 82 patients. The full model improved concordance and response AUROC over a clinical baseline but was not formally compared with the strongest constituent model [49]. Nevertheless, the absence of consistent superiority is not evidence of no fusion benefit.

Spatially resolved omics and marrow-niche mapping

Beyond diagnosis, the same computational approaches have opened adjacent frontiers. Because bulk RNA sequencing destroys the marrow architecture by homogenizing tissues, multimodal AI is being paired with spatially resolved multi-omics data, including sequential immunofluorescence, imaging mass cytometry, and high-plex image analysis data, to recover location-specific biology from archival specimens. Unbiased spatial clustering has revealed previously unrecognized tertiary-lymphoid-structure–like aggregates within the AML marrow; discovered in a small number of patients, the signature of these aggregates was validated as outcome-predictive in an independent 480-patient transcriptomic cohort [50]. A parallel myeloma study mapped the trephine microenvironment (MoSaicNet/AwareNet; tissue- and cellular-classification AUROC > 0.98), linking the spatial organization of immune and tumor cells to disease, although in fewer than 20 patients [51]. Serving discovery rather than routine diagnosis, the spatial biomarkers this approach surfaces are precisely the signals that a mature multimodal diagnostic system would ultimately need to ingest.

The current state of genuine integration

Using the operational definitions provided in Sect. "Review search strategy and evidence classification", this review identified no peer-reviewed, prospectively validated marrow AI systems that jointly ingested matched WSI or smear data, raw flow cytometry events, cytogenetic or molecular outputs, and structured clinical metadata to yield a unified, WHO-HAEM5/ICC-aligned recommendation. The existing systems are strong single-modality tools, compelling cross-modal predictors, a few disease-specific partial-fusion systems with mixed, endpoint-dependent findings, and general pathology foundation models that align slides with reports and gene expression. Notably, even the reference laboratory most closely associated with integrated diagnosis, the Munich Leukemia Laboratory, which routinely issues combined morphology–flow–cytogenetics–molecular reports [52], has pursued machine learning in a genomics-first manner, classifying entities from whole-transcriptome sequencing rather than fusing morphological data with other modalities [53]. In parallel, large language models are beginning to extract and structure the free text of pathology and BM reports [54], a step toward the case-level synthesis integrated AI will ultimately require. Therefore, the realistic future is best conceived not as one monolithic foundation model but as a case-level orchestrator around multiple specialist encoders, matching both the current workflow and the structure of the available evidence (Fig. 1c, Table 1).

Table 1.

Representative studies across the maturity spectrum of AI systems for bone marrow examinations

Study (yr) [ref]	Evidence class	Inputs → label	Cohort & validation	Key result	Key consideration
Matek et al. (2021) [12]	Single modality	Single cells → 21 morphologic classes	171,374 images/945 patients; single-lab benchmark	Field-defining public benchmark; expert-level average accuracy	Cell labels ≠ specimen-level diagnosis; rare-class precision very low
MarrowQuant 2.0 (2023) [24]	Single modality (histology)	Trephine WSI → cellularity & compartments	250 specimens; vs. hematopathologist consensus	Cellularity intraclass correlation coefficient: 0.98; open-source (QuPath)	Cautioned against MPN/megakaryocyte hyperplasia; a quantifier, not a diagnostic classifier
Tayebi et al. (2022) [6]	Single modality	Aspirate WSI → cell detection + HCT summary	Single-institution; retrospective	ROI accuracy: 0.97; cell mAP: 0.75; F1: 0.78	External validation not feasible; not comprehensive diagnosis
Scopio FF-BMA (2024) [21, 22]	Single modality (device)	Aspirate WSI → quality/blast/plasma-cell/M:E (assistive)	795 specimens/3 sites; multicenter, retrospective	1,385 cells/specimen; 87.0% of fields of view accepted; agreement: 91.1%; FDA De Novo (new Class II)	Assistive only; blast specificity 84.8% → 73.1% in MDS/MPN
Lewis et al. (2024) [30]	Single modality + cross-modal	Raw flow → diagnosis (uni); → 9 cytogenetic/32 variant labels (cross-modal)	1,820 samples (736 AL); single-center	Dx AUROC: 0.96; abstention lifts accuracy: 91.4% → 97.9%; genetics strong only for a few (PML::RARA 0.93 > NPM1 0.81)	Panel-dependent; genetic prediction is cross-modal, not fusion
Cheng et al. (2024) [31]	Single modality	EuroFlow ALOT flow → acute leukemia screen	241 patients; single-center	AML accuracy: 94.6% (sensitivity: 90.0%); B-ALL: 98.2% (sensitivity: 90.9%)	Flow-panel dependence limits generalization; weak for MDS
IPSS-M (2022) [36]	Molecular–clinical (prognostic)	Somatic mutations + cytogenetics + clinical → prognostic risk	2,957 patients; externally validated (754 J-MDS)	Re-stratified: 46.0%; C-index improvement ~0.05 vs. IPSS-R; handles missing genotype (best/avg/worst)	Integrates molecular, cytogenetic, and clinical variables, not imaging/flow modalities; next-generation sequencing access equity limits reach
Eckardt et al. (2022) [43]	Cross-modal	Smear image → AML dx & NPM1 (molecular) label	1,251 AML/236 controls; single-center, internal only	AML AUROC: 0.97; NPM1: 0.92 (image-only)	Single-institution; image-only; manual region selection (bias source)
Kockwelp et al. (2024) [44]	Cross-modal	Whole-smear scans → therapy-relevant genetics	408 AML (+71 temporal); single-center	CBFB::MYH11 ~ 0.91; but NPM1 ~ 0.70, MRC ~0.68, ELN-favorable ~0.64	AML-specific; limited genetic signal in morphology; not fused at inference
Yu et al. (2025) [47]	Partial fusion	AI-derived biopsy pathology features + clinical → MPN subtype	Internal + external; MPN-specific	Macro-AUROC: 0.96 for both image-only and comprehensive models; average precision: 0.75 → 0.81; external image-only macro-AUROC: 0.94	Only image model externally validated; 2016 WHO labels; no flow or genomics
Wang et al. (2025) [48]	Partial fusion (late fusion)	WSI + clinical → MPN dx/subtype (score-level combination)	Multicenter; MPN-specific	Fusion significantly improved overall MPN (0.93 vs. 0.89 DL/0.85 clinical) and ET (0.89 vs. 0.84/0.83); not consistently across subtypes	Retrospective; no raw flow/molecular; best modality is disease-dependent (clinical alone matches fusion in PV)
ImmunoCast-MM (2026) [49]	Partial fusion	Cytomorphology +18F-FDG PET/CT (+ clinical covariates for outcomes) → immune decoding/response	243 newly-dx MM; single-center, retrospective	Decoded indices vs. flow ρ: 0.68–0.81; C-index: 0.58 → 0.75; response AUROC: 0.55 → 0.81 (vs. clinical baseline)	Flow used for alignment/validation (82/243), not an input; no formal comparison with strongest constituent model; not comprehensive
mSTAR (2025) [41]	Fusion template (off-marrow)	Multimodal pre-training (WSI + report + gene expression); WSI at inference → many tasks (+ gene expression for fusion)	32 TCGA cancer types (including diffuse large B-cell lymphoma); not marrow-specific	Multimodal beats vision-only across 97 tasks (+1.8% in fusion C-index and 2.6% in mutation macro-AUROC over the next-best model)	Mostly solid tumors; bone marrow essentially unrepresented; tri-modal fusion is a pre-training signal

The results are illustrative and not directly comparable because the studies differ in task, cohort, reference standard, and validation design. The "Inputs → label" column distinguishes cross-modal prediction from fusion, and evidence classes follow the inference-input definitions in Sect. "Review search strategy and evidence classification"; molecular–clinical prognostic integration and mSTAR are included as context. No comprehensive marrow system meeting the stated definition was identified in the reviewed literature. Direct comparator evidence for partial fusion was sparse and endpoint-dependent; nevertheless, lack of consistent superiority does not imply the absence of any benefits

Technical foundations and the data ecosystem

Five design choices

Most AI systems used for BM examinations are based on five choices: representation learning, instance aggregation, cross-modal fusion, explainability, and uncertainty management. Cytology models use detector–classifier or direct classifiers; WSI systems rely on patch or cell embeddings with multiple-instance learning; flow cytometry systems treat each event as an instance in a bag with attention or set-based mechanisms; and multimodal systems add structured clinical or genomic features. Pre-training is another decisive axis, with domain-specific foundation models increasingly displacing the generic ImageNet backbones.

Multilevel explainability

Explainability in marrow AI should be understood as multilevel rather than as a single saliency map: cell exemplars for cytology, patch heatmaps for biopsy WSI, population attention for flow cytometry, and feature-attribution for structured clinical or genomic variables. AML image-genomics studies have used occlusion maps to reveal morphological correlates of the NPM1 mutation [43]. For adoption, the progression from an uninterpretable “red heatmap” to a retrievable evidence packet linking a prediction to specific cells, regions, populations, and variants is likely crucial for physician trust and medico-legal defensibility.

Uncertainty quantification and abstention

Uncertainty quantification remains underdeveloped in marrow-specific work; however, it is especially important because the most difficult cases, such as those involving suboptimal aspirates, patchy infiltrates, rare malignant cells, dry taps, and flow-panel drift, are precisely where overconfidence is dangerous. Studies should report calibration, confidence intervals, class-wise error structures, and a prespecified abstention policy; ideally, they should employ ensemble or conformal methods and prospectively evaluate failure modes. A human-in-the-loop workflow triages low-confidence cases to an expert so that overconfident predictions of ambiguous materials do not propagate into clinical decisions. Partial precedents exist: an attention-based flow cytometry model whose abstention on low-confidence cases increased the accuracy from 91.4% to 97.9% [30], and a pipeline that flagged ambiguous acute leukemia cases for manual review [29]. However, such calibrated abstention remains the exception rather than a built-in default.

Datasets, benchmarks, and the fragmentation problem

The public data ecosystem is fragmented and strong for single-modality data: the Matek corpus anchors smear work; named public resources now serve specific niches—BMASID for aspirate-smear cell detection (230 annotated smears) [55], the MK-11 megakaryocyte dataset (7,204 images across 11 subtypes) [56], and a pediatric-leukemia aspirate-WSI collection (246 cases, 49 consensus cell classes, though its documentation notes that annotation-session labels are flawed and consensus labels should be used) [57]—while FlowCAP and HDCytoData underpin flow-cytometry benchmarking [58, 59], and cohorts such as TCGA-LAML supply the molecular side. Widely adopted resources do not provide the pairing that integration requires, i.e., digitized marrow morphology, raw flow cytometry, molecular profiling, and longitudinal outcomes for the same patients, at scale. Researchers should also distinguish benchmarking from development and clinical validation datasets because institutionally governed paired datasets are necessary for clinically meaningful fusion research.

Bottlenecks impeding clinical adoption

Despite its promise, moving multimodal AI from retrospective in silico validation to prospective clinical practice entails severe technical, data, ethical, and regulatory bottlenecks (Fig. 1d).

The generalization crisis and batch effects

The most critical vulnerability of deep learning in computational pathology is its poor robustness across heterogeneous clinical environments. Models are hypersensitive to pre-analytical variations such as smear preparation, staining, section thickness, scanner optics, and image compression, which induce batch effects (site-specific artifacts learned as if biological). Therefore, the strong internal discrimination can degrade, sometimes to clinically inadequate levels, at an institution with different scanners or normalization protocols. Because the marrow class distribution is long-tailed and rare morphologies are decisive, steeper declines in external validation can be expected than in standardized solid tumor tasks, unless harmonization, color normalization, cross-site validation, and uncertainty-aware abstention are designed from the outset. This risk is not hypothetical; a deployed flow cytometry AML classifier showed lower sensitivity in real-world use than in validation (69.0% vs. 77.0%) and, separately, a transient decline in the F1 score was associated with laser failures in two cytometers and a reduced CD45 signal [33]. Federated learning—training across decentralized datasets without centralizing sensitive data—is a leading remedy and has been demonstrated for gigapixel computational pathology [60] and extended to cross-institutional modality heterogeneity faced by marrow fusion [61].

Heterogeneous fusion and missing modalities

Genuine fusion must integrate image, tabular, and sequence data that is produced by different laboratories on different timescales and is routinely incomplete. Incomplete flow panels, absent cytogenetics, and misaligned tests are the norm. Therefore, robust systems must explicitly handle modality missingness through architectures that degrade gracefully, imputation, or modality-dropout training. Hematology offers a precedent: the molecular IPSS reports the best-, average-, and worst-case risks when a genotype is only partially available, converting missing data into a bounded range rather than an outright failure [36]. Managing this heterogeneity, not incremental architectural novelty, is the defining scientific problem of marrow multimodal AI.

Data scarcity, class imbalance, and nomenclature

Datasets are small relative to task complexity; class imbalance is extreme for decisive rare cells (mast cells, basophils, megakaryocyte maturation stages, mitotic figures) and rare diseases; annotation is costly; and the ground truth itself carries inter-observer variability, because expert morphology shows only substantial—not perfect—agreement. A deeper instability is guideline-level: WHO-HAEM5 does not specify a minimum blast threshold for AML with mutated NPM1, whereas ICC 2022 requires ≥ 10% blasts; consequently, cases below this threshold may be classified differently across systems and require careful clinicopathologic correlation [2–4]. Compounding these issues, staining conventions (MGG/Pappenheim vs. Wright-Giemsa), scanner and cytometer variability, and divergent classification schemes leave no harmonized labeling standards across sites.

Interpretability, regulation, computation, and equity

The black-box nature of heavily layered multimodal transformers obscures the features or gene variants that drive classification, complicating the responsibility that clinicians must retain. By June 2025, the FDA had authorized 20 AI/ML-enabled devices in the hematology category; however, none were cleared for standalone, autonomous diagnosis of acute leukemia, reflecting appropriate caution for high-risk malignancies [62]. Regulatory frameworks designed for statically “locked” algorithms are only beginning to address continuous-learning models, for which predetermined change-control plans have been proposed [63]. Professional bodies require formal validation of WSI before diagnostic use [64], and the European Union now explicitly regulates high-risk medical AI [65]. Another concern is access, not accuracy: the computational and energy footprints of large transformers and their dependence on costly graphics processing unit (GPU) infrastructure risk excluding resource-limited centers from precision oncology access.

Governance and medico-legal risks

Whole-slide marrow scans are large and can contain identifying artifacts, while raw FCS files, variant calls, and longitudinal records add governance complexity, and paired cohorts are often too small or disease-specific to share openly. Therefore, federated, privacy-preserving, or controlled-access paradigms are more realistic than unrestricted open data sharing. Medico-legal risk is another unresolved topic; residual uncertainty is a genuine concern in difficult cases, such as those involving atypical morphology, treatment effect, fibrosis, scant material, or unusual genotype–phenotype combinations, where overconfident AI is especially dangerous, reinforcing the assistive, human-confirmed role that current devices and guidelines already mandate.

Future directions and research priorities

The central gap: paired, procedure-faithful case assemblies

The central scientific gap is not simply “more data” but the absence of paired, procedure-faithful case assemblies in which aspirate cytology, biopsy histology, flow cytometry, cytogenetics, molecular testing, and curated metadata are aligned at the case level and linked to time-aware endpoints. Without these, most AI systems for marrow examinations will remain modality-specific or conflate surrogate tasks with the genuine objective of an integrated diagnosis. Building these cohorts—multi-institutional, governed, and often federated—is the highest-priority infrastructure investment in the field.

Staged prospective validation

A robust evaluation program should proceed in stages rather than directly advancing to a clinical intervention (Table 2). The stages are as follows: multicenter retrospective external validation with frozen algorithms and temporal holdout; a silent prospective study in which the AI runs in real time but does not affect care, so calibration, failure modes, and discordance with final sign-out are measured live; and a live clinical-impact study, ideally stepped-wedge or cluster-randomized, measuring turnaround, re-review burden, diagnostic concordance, and safety events. A rigorous design will use case-level reference standards from a multidisciplinary consensus after all routine modalities are available rather than employing single-modality labels and will permit the AI to abstain. Stratified analyses across specimen quality, disease category, treatment status, platform, and site should be mandatory, because these choices, rather than minor architectural differences, determine whether a study answers a clinically meaningful question.

Table 2.

Staged validation framework for comprehensive AI analysis of the bone marrow

Validation stage	Recommended design	Primary endpoints	Critical secondary endpoints	Applicable reporting/appraisal frameworks	Current status in AI assessment of bone marrow
Preclinical external validation	Multicenter retrospective, temporal holdout	AUROC/macro-F1/AP by task	Calibration, subgroup performance, abstention	Align, where applicable, with prediction-model reporting and appraisal (TRIPOD + AI; PROBAST + AI), diagnostic-accuracy reporting (STARD-AI), and imaging AI reporting (CLAIM); emphasize data provenance, site-wise splits, calibration, and inputs available at deployment	Reached by several multicenter retrospective studies (digital cytology, flow); temporal/prospective external validation still uncommon
Silent prospective evaluation	Real-time shadow mode in routine workflow	Concordance with final consensus diagnosis	Turnaround latency, failure capture, drift, alert burden	DECIDE-AI (reporting for early live/shadow-to-live evaluation): human–AI interaction, failure capture, usability	Rarely attempted; isolated prospective validations exist, but shadow-mode evaluation in a live workflow is essentially unpublished for marrow
Clinical-impact evaluation	Stepped-wedge, cluster trial, or controlled before–after	Time to actionable result; change in accuracy or test prioritization	User trust, override rate, downstream testing, safety events	Trial standards: SPIRIT-AI (protocol) and CONSORT-AI (report); report subgroup error analysis and decision thresholds	No published clinical-impact trial of marrow AI—the principal evidence gap
Post-deployment surveillance	Continuous monitoring	Calibration drift, false-reassurance rate	Revalidation triggers, version-change effects	Change-management documentation aligned with the FDA predetermined change-control plan (PCCP) guidance (regulatory)	Isolated real-world deployments with drift monitoring; no standardized surveillance reported

The reporting and appraisal frameworks listed here are voluntary community standards rather than mandatory requirements, and the appropriate framework depends on the study's design. The FDA PCCP document is a regulatory agency guidance rather than a voluntary reporting framework

Strengthening validation and reporting standards

A recurring weakness of the literature describing multimodal AI systems for BM examinations is that the evidence base is under-validated and inconsistently reported; most studies are single-center and retrospective, and the adoption of standardized reporting frameworks is low. Two reporting elements are particularly decisive because they distinguish genuine fusion from relabeled cross-modal prediction: the exact modalities available to the model at inference and the framework used to derive the labels. A system that consumes only a digitized smear yet reports a genetic endpoint is performing cross-modal inference, not fusion; PROBAST + AI operationalizes this test, asking whether “the predictors included in the model [were] available at the time the model was intended to be used” [66]. Studies should also report the intended use, specimen source and quality controls, staining and scanner details, flow-panel composition, site-wise and temporal data splits, calibration, decision thresholds, missing-data handling, and reader-comparator design, aligning with the established framework that matches their design: TRIPOD + AI and its companion PROBAST + AI for prediction models [66, 67], STARD-AI for diagnostic accuracy [68], CLAIM for imaging AI [69], DECIDE-AI for early live evaluation [70], and SPIRIT-AI and CONSORT-AI for prospective trials [71, 72]. For deploying laboratories, this means identifying where AI can genuinely help—triage, pre-classification, quantitative assistance, discrepancy detection, or report standardization—and validating the AI tool against the laboratory's own preparation, staining, scanner, panel, and patient mix with drift monitoring and operator-confirmed assistive deployment.

Limitations of this review

This non-systematic narrative review was conducted by a single author without independent duplicate study selection or formal risk-of-bias assessment. Therefore, study selection and interpretation may be susceptible to selection and confirmation biases. Heterogeneity in tasks, comparators, and validation designs limited direct cross-study comparisons. Table 1 should be regarded as illustrative rather than exhaustive. The conclusions are bound by the evidence available at the time of manuscript preparation.

Conclusions

BM examination is at the threshold of technological transformation; however, the evidence must be read with precision. The successes of AI are real but partial—convincing within individual modalities and promising in narrow cross-modal and partial-fusion settings. Within the literature reviewed for this manuscript, no comprehensive, prospectively validated system was found that jointly ingests morphology, flow, and molecular data and clinical metadata to provide a unified, guideline-aligned diagnosis.

Overall, integrated multimodal AI for BM examinations has entered a promising but early translational phase, with a clear path to consolidation. Consolidation will require large, multi-institutional, federated paired datasets that curb batch effects and enable rare-disease modeling; fusion architectures that handle missing and asynchronous modalities; transparent, calibrated, abstention-aware outputs; harmonized nomenclature with public multimodal benchmarks; and staged prospective validation against a multidisciplinary, case-level reference standard rather than surrogate labels. The decisive threshold, which has been reported in several solid-tumor computational pathology settings but not yet established for the BM, is a rigorous demonstration that fusion can improve on the performance of the strongest single-modality comparator. Current marrow comparisons are sparse and endpoint-dependent: gains have been reported for selected endpoints, but consistent superiority across tasks has not yet been demonstrated. Nevertheless, this limited evidence does not imply the absence of benefits. Until then, integrated multimodal marrow AI should be advanced with disciplined ambitions, pursued deliberately, and reported transparently.

Author’s contributions

The sole author (Seung Yeob Lee) conceived the review, performed the literature synthesis, and wrote and approved the final manuscript.

Funding

This work was supported by the Korean Cell-Based Artificial Blood Project funded by the Korean government (The Ministry of Science and ICT; The Ministry of Trade, Industry and Energy; The Ministry of Health and Welfare; The Ministry of Food and Drug Safety; Grant No. RS-2023-KH140972); the National Research Foundation of Korea (Grant No. NRF-2022R1C1C1011518); the Bio & Medical Technology Development Program of the NRF funded by the Korean government (MSIT; Grant No. RS-2023-00236157); and the Biomedical Research Institute fund of Jeonbuk National University Hospital.

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

Not applicable. This review was based solely on previously published literature and does not include new human participants, human specimens, identifiable human data, or animal experiments.

Consent for publication

Not applicable.

Competing interests

Seung Yeob Lee was the Assistant Editor of Blood Research while conducting this study. Editorial Board Member status has no bearing on editorial considerations. In addition, Seung Yeob Lee is the founder of DiaPLT Co., Ltd; this interest is not related to the content of this manuscript.