Artificial Intelligence in Bone Marrow Histological Diagnostics: Potential Applications and Challenges.

Introduction

The complexity of the bone marrow (BM), consisting of many different cell lineages in various stages of differentiation, which dynamically respond to local or systemic signals, makes it a challenging organ for diagnostics. Basic histopathological evaluation of a BM biopsy (BMB) comprises the assessment of different lineages for age-related abundance (hypo- or hyperplasia), effective maturation and normal topography, abnormal cellular features (dysplasia), and the presence of uncommon elements such as lymphoid aggregates, fibrosis, amyloid, or granulomas. This initial evaluation is generally performed on a hematoxylin and eosin (H&E)-stained glass slide, supported by additional histochemical stains (PAS, Giemsa, Perls; reticulin) for the assessment of specific cell features and lineages and stroma components. Additional immunohistochemical stains or molecular tests are added for more specific information, depending on the initial assessment, the clinical setting, and differential diagnoses that have to be excluded.

Since the initial evaluation is essential in guiding further diagnostic tests, it requires specific diagnostic experience. The current ongoing digitalization of pathology and the future introduction of artificial intelligence (AI) in clinical BM diagnostics are expected to have a significant impact on routine diagnostics. Possible applications range from providing a quantitative summary of the slide (e.g., cell counts) to a global qualitative impression of a whole-slide image (WSI) to help establish whether there is disease, which cell lineages are involved, and which group of diseases has to be considered, or to suggest a specific diagnosis, including prognostic and/or predictive factors. In this perspective, we introduce basic concepts of AI, envision how AI could support routine BM histological diagnostics, show first results, and discuss the obstacles that lie ahead.

AI in Pathology

A Brief Introduction on AI

AI is a broad concept, relating to systems (machines or computers) that are capable of fulfilling tasks at a human level of performance [ 1 ]. While AI is often used interchangeably with related terms such as “machine learning” (ML) and “deep learning” (DL), there are meaningful distinctions. ML is a subfield of AI, consisting of computer algorithms that are capable of learning tasks from data, without being explicitly programmed [ 2 ]. In essence, these algorithms do not operate on pre-programmed rules, but are allowed to decide on rules themselves by finding patterns in data. DL is a subfield of ML which uses deep neural networks of multiple layers of artificial neurons that learn their own level of representation of the data by repeatedly updating the parameters of the neurons in a process called “training” [ 3 ].

The vast majority of recent research papers published on AI in medical image analysis uses some form of DL [ 4 ]. This trend is caused by the fact that DL techniques generally have a high accuracy and ease of implementation [ 4 ]. With the introduction of whole-slide scanners in the 1990s [ 5 ] and the current expanding digitalization of routine diagnostic histological slides [ 6 ], the number of papers applying DL to histopathology has been growing annually [ 5 ].

Different Learning for Different Tasks

In a recent perspective, Kann et al. [ 7 ] argued that AI in a medical context “can be viewed as a spectrum of algorithms, the utility of which is inextricably linked to the characteristics of the task under investigation.” Indeed, different categories of ML algorithms may suit different tasks that are encountered in histopathology. In Figure 1 , we summarize the main categories of ML algorithms and examples of tasks that can be solved by applying these to WSIs. We present these categories as a gradient, arranged according to how numerous or “fine-grained” the annotations that are necessary to train the network.

Figure 1a shows an example of (fully) supervised learning: all elements related to the task the algorithm is trained for are annotated. In less strongly supervised methods of semi-supervised learning (Fig. 1b ), only a subset of images/elements within the images is fully annotated. In weakly supervised learning (Fig. 1c ), only a single label is used for the entire WSI (e.g., a diagnosis such as acute myeloid leukemia). Lastly, Figure 1d shows an example of unsupervised learning. Instead of annotated data, unsupervised algorithms cluster images into groups that share characteristics. Because the learning happens without annotations, it is often difficult to readily understand the difference between the clusters [ 5 ].

The choice of category of ML to apply to a specific task is a non-trivial decision; there are often different solutions for the same task, and frequently different algorithms are combined to perform a task. While supervised algorithms typically outperform less supervised methods due to the fact that the algorithms learn from stronger knowledge, they do so at the cost of fine-grain annotations, which are time consuming and require expertise. An example is the comparison of the performance of an (fully supervised) algorithm that was trained to recognize megakaryocytes by pixel-wise manual annotation of single cells, as depicted in Figure 1a , and the recognition of megakaryocytes after only marking a cell as megakaryocyte with a dot (i.e., semi-supervised learning), followed by extrapolation with the mean cell size of the megakaryocytes in the training set, as shown in Figure 2 . In this example, the fully supervised algorithm produces segmentation that more closely matches the cellular boundaries. The rough cell contours that result from the semi-supervised approach can be smoothened by applying additional algorithms that recognize cell borders, which makes the prediction more precise.

To prevent extensive annotations, semi-supervised/weakly supervised methods can be trained with slide-level annotations, which could be automatically extracted, e.g., from electronic pathology reports, although at the cost of lower performance. The gap between fully supervised methods, with manually annotated small tissue areas in limited cases, and semi-supervised and weakly supervised methods has been gradually decreasing, with some papers approaching the performance of fully supervised methods [ 9 , 10 , 11 ]. This, however, requires huge datasets of tens of thousands of WSIs for the training and is therefore mainly used for common tumor types [ 9 ]. The application of free text recognition programs, which can automatically extract diagnostic information from electronic reports, is a potentially important tool to be able to collect these large datasets [ 11 ].

Applications of AI to BM Histology

In the following sections, we describe the current developments in the field of AI application for BM histology.

Identifying BM Cells and Structures

Automated recognition of tissue structures and cell lineages (an example is shown in Fig. 1a, b ) in BMB WSIs allows for quantification of overall cellularity, marrow distribution, and the abundance and maturation of the different hematopoietic cell lineages with direct visual control. These data would provide a global impression of the WSI and could also be used to support disease recognition and classification.

While automated cell counting of peripheral blood is fully integrated in routine laboratories, cytomorphology of blood and BM smears for hematological neoplasms still requires examination by experienced technicians and hematologists. To support these diagnostics, during the past few years experimental algorithms for cell identification in digitized blood smears [ 12 ] and automated selection of suitable areas and single-cell recognition of BM cells in aspirate smear WSIs were developed [ 13 ]. Applications of AI to blood and BM aspirate smears have been discussed elsewhere [ 14 ].

The evaluation of histology is, however, much more complex due to tissue architecture and the variation in object shape due to cutting effects. First attempts using ML have been made to identify and enumerate a single-cell lineage that was stained with specific antibodies, e.g., plasma cells (CD138) [ 15 , 16 ], or lymphoid and plasma cells (CD3, CD20, CD138, MUM1) [ 17 ]. More advanced studies used the morphological features in H&E-stained WSIs, without the addition of immunohistochemical stains, to discriminate between megakaryocytes, erythroid, and myeloid cells in patients with myeloproliferative disease [ 18 , 19 , 20 ]. There are also studies that segmented hematopoiesis and fat tissue for cellularity quantification [ 21 , 22 ]. We recently extended the recognition of megakaryocytes, erythroid, and myeloid cells with that of fat cells and bone tissue in PAS-stained slides of normal BMBs, providing a comprehensive evaluation of the tissue (Fig. 2 ) [ 8 ].

Indeed, AI-supported visualization or overlays of different cells and structures on the WSI could help the pathologist to focus attention, detect abnormalities, and aid in further diagnostics. We show a hypothetical example of lymphoma detection in an H&E-stained BMB WSI in Figure 3 . Clinical studies on AI-assisted reading of routinely scanned H&E slides in BM pathology have not yet been performed, although first progress is being made in other areas of pathology [ 23 , 24 , 25 ].

Recognizing Disease

A more advanced diagnostic approach is the application of DL algorithms to recognize a broad variety of specific myeloid and/or lymphoid disease categories in WSIs of BMB. Such broad diagnostic studies have yet to be performed. A more limited but clinically relevant task would, e.g., be to recognize dysplasia, which has been studied on BM aspirates and peripheral blood smears in the context of myelodysplastic syndromes (MDS) [ 26 , 27 ]. Even more challenging would be to discriminate MDS from reactive dysplasia in BMB (Fig. 1c ).

A first study on applying ML to MDS used tissue microarrays made from BMBs [ 28 ]. The authors extracted morphological features from the H&E-stained tissue microarrays and were able to train regression models using these features, which could accurately predict for (cyto)genetic aberrations and distinguish MDS and MDS/myeloproliferative neoplasm (MPN). Moreover, when the authors performed unsupervised clustering of the H&E-extracted morphological features, they found clusters of myelodysplastic subgroups distinct from the normal controls. In addition, when analyzing the interpretability of their models, the authors found that the predictions of their regression models correlated with automatically measured tissue features such as average nucleus size and eccentricity and H&E stain densities. This study gives a glance at the strength of multi-level image analysis for recognizing disease in BM diagnostics. It should be noted, however, that all of the results were reported without an independent set of cases (a test set), by using a small cohort of only 194 MDS or MPN/MDS diagnosed patients and 11 control patients, which may all significantly skew results.

Disease-Specific Algorithms

As the aforementioned study by Brück et al. [ 28 ] showed, ML can be applied within a disease-specific context, after the differential diagnosis was already narrowed down, to automate specific tasks while viewing the case. While this is useful in a research setting when a large series of specific patients is collected, it may be less helpful or even distracting in a clinical setting if the initial diagnosis was not correct.

For disease-specific ML, one of the most attractive study subjects in BMBs is MPN. For the discrimination of the separate disease entities essential thrombocytopenia (ET), polycythemia vera, and primary myelofibrosis (PMF), BM histology is a major diagnostic criterion and histological differences are largely determined by abnormal megakaryocyte morphology and distribution [ 29 ], the interpretation of which is subjective [ 30 , 31 ]. Sirinukunwattana et al. [ 20 ] developed a ML algorithm for automated identification, quantitative analysis, topography, and cytomorphological analysis of megakaryocytes. This indeed allowed discrimination of MPN in general and ET from control BMB, based on a single (megakaryocytic)-cell lineage. Adding information of the neighboring megakaryocytes and overall distribution to the algorithm allowed further classification between ET, polycythemia vera, and PMF in a follow-up study [ 32 ].

Another question that could potentially be addressed by ML is the grading of BM fibrosis in MPN. Significant fibrosis at disease onset or during disease progression is associated with a worse prognosis [ 33 ]. Fibrosis grading that is based on different protocols of fiber stains tends to suffer from both technical and interobserver variations [ 34 ]. Digital image analysis of special fiber stains has been proposed for quantification, but is not applicable for routine diagnostics [ 35 ]. In a recent study, Ryou et al. [ 36 ] used reticulin fiber-stained WSIs to develop algorithms to visualize and grade areas with increased fibers in WSIs of MPN and showed good correlation with morphological grading and MPN type. Adding megakaryocytic features as described in their previous study [ 20 ] further improved the diagnosis and allowed discrimination of ET and prefibrotic PMF cases. The latter were characterized by areas of microfoci of fibrosis. To circumvent the use of additional stains, we made a first attempt in H&E-stained BMB slides of MPN patients to automatically detect clinically significant fibrosis (WHO grade 2 and 3). Figure 4 gives a summary of the results.

Irshaid et al. [ 37 ] tackled a different disease-specific application with the ML-supported detection of blastic transformation in the BMB of patients with follicular lymphoma or chronic lymphocytic leukemia. The authors trained several algorithms to quantify small cell and large cell lymphoid infiltration, followed by a training with the outcome of the pathological diagnosis of transformation that was based on quantification of morphological features of atypia, blastic transformation, fibrosis, and proliferation index measured by Ki-67. The results were compared to the final clinical decision of transformation that was based on lymph node biopsy results and/or clinical features of disease progression requiring aggressive chemotherapy. DL algorithms were able to augment the pathologist morphological scoring of transformation, and results were more reproducible.

Phenotype-Genotype Correlation

The continuous search in pathology for markers that are either disease-specific, hold prognostic information, or are predictive for an evolving treatment arsenal generally requires large series of patients with good clinicopathological correlation. While many traditional markers consist of morphological features or specific protein expression patterns, the past years showed an explosion of genetic markers as a result of the availability of next-generation sequencing (NGS). Although specific genetic changes may result in unique diseases, like the BCR :: ABL1 translocation in chronic myeloid leukemia, and specific pathological features can reveal an underlying genetic event within a disease category, like the cup-like nuclei in AML with NPM1 and FLT3 mutations [ 38 , 39 ], most genetic drivers are impossible to detect by traditional pathology. First results of AI predicting genetic changes in hematologic diseases, which also can be visually recognized in smears by experienced hematopathologists, are an encouraging step and could support diagnosis in situations when expertise and/or genetic testing is lacking. Eckardt et al. [ 40 , 41 ] trained an algorithm to discriminate BM smears of AML patients with a translocation (15; 17) or a NPM1 mutation from other AML types or normal controls.

The recognition of genetic changes in the far more complex tissue sections, especially if they are impossible to detect by morphology, is a further step to support diagnostics. In a proof-of-concept study, we were able to predict MYC translocation, as determined by FISH, with a high sensitivity (0.93) in H&E WSIs of biopsies with diffuse large B-cell lymphoma, albeit with lower specificity (0.52) [ 42 ]. In the study of Sirinukunwattana et al. [ 20 ], megakaryocyte morphology and distribution not only distinguished different subtypes of MPN, but also showed a correlation with the presence of a mutation in one of the driver genes JAK2, CALR , or MPL . Although these driver mutations correlate with some general clinical and morphological characteristics in larger MPN cohorts [ 43 ], they cannot be reliably predicted for individual patients based on clinico-morphological correlation. In the aforementioned study on MDS and MDS/MPN, Brück et al. [ 28 ] extracted dozens of qualitative and quantitative morphological features from the BM images. These data were combined for extensive supervised and unsupervised data analysis with a large array of clinical characteristics, laboratory values, and genetic information from NGS and cytogenetic tests. Several algorithms were trained to identify correlations between morphology and other features, including the chance of specific genetic aberrations. Indeed, several mutations and cytogenetic aberrancies, such as deletion 5q with its specific megakaryocyte morphology, could be predicted with a high accuracy.

Future Perspectives

Most aforementioned examples are based on routine H&E WSIs, with an occasional example that uses immunohistochemical stains. Increasingly, more complex research data are being created by in situ imaging technologies such as multiplex immunohistochemistry or immunofluorescence and mass spectrometry. These technologies add information on protein or gene expression of individual cells to the morphological features and spatial distribution. The high-dimensional nature and breadth of these data make them very suitable for ML, to discriminate between target cells, bystanders, and stromal structures, to provide quantitative data, and to describe spatial distribution of specific populations [ 44 ]. Such multiplex tissue imaging techniques are currently being developed for decalcified, FFPE tissue slides in BM research [ 45 ]. Initial studies have analyzed the hematopoietic niche [ 46 ], and the microenvironment in chronic myeloid leukemia, AML, B-cell acute lymphoblastic leukemia [ 47 , 48 ], and multiple myeloma [ 49 ].

Several other developments in the field of AI in BM pathology can be anticipated. Algorithms that give a quick, global assessment of the WSI would be very supportive to guide differential diagnoses and additional tests, both in situations of limited resources and/or lack of expertise, and in complex diseases requiring integration of additional tests. Automated integration of high-quality morphology, and clinical and laboratory data will boost the precision of the prediction. Another relevant development for digital pathology from the ML field is out-of-distribution detection [ 50 ]. Current ML methods do not allow for recognition of cell types or diseases that the algorithm has not been trained for. Out-of-distribution methods allow an algorithm to determine when newly presented data are far outside (in terms of appearance) the set of images it was trained with, thus giving algorithms the ability to say “I don't know.” Such methods would add a great deal of robustness and reliability if integrated into ML systems designed for clinical use [ 50 ].

In specific diseases, prognostic and predictive information can probably be refined by adding information derived from the WSI by AI, although extensive clinical validation is required. Overall, the future in hematopathology is expected to exploit AI-assisted automated collection and integration of data derived from several sources, including cyto- and histomorphology, clinical data, laboratory data, and genetics to support diagnosis [ 51 ]. It is important to emphasize the role of the hematopathologists in this process, as they will be essential for quality control, for clinicopathological correlation and in making new connections and asking new questions to further advance diagnostics.

Challenges and Considerations for the Introduction of AI in Pathology

There is a large number of challenges that must be overcome before AI can be widely introduced into the routine clinical pathology workflow. Indeed, despite the large number of papers on AI applied to pathology [ 5 ] and medical imaging in general [ 4 ], only a handful of FDA- or EMA-approved DL applications exist [ 7 ].

One important fact that currently limits the applicability of AI is that most research papers train their ML methods on data that are unrepresentative of data found in clinical practice [ 52 ]. Aspects such as tissue preparation, slide staining, and tissue/slide artifacts can introduce a large variation in WSIs, especially between different clinical centers. This variation, when unaccounted for in the data used to train the algorithm, may have a significant negative impact on its performance [ 53 ]. Therefore, it is imperative that during algorithm development, training data are selected to capture as much variation as possible. A research area of medical ML that has received little attention is the question of how the necessary amount of variation (e.g., necessary number of different pathology laboratories producing the training data) differs depending on the complexity of the clinical question [ 52 ].

Another crucial condition of clinical adoption of AI is extensive validation to prove its usefulness and identify weaknesses/biases [ 7 , 52 ]. While most AI papers use an independent test set to validate performance on data the algorithm has not yet seen, this is frequently drawn from the same (internal) data source [ 52 ]. Validation on external test sets gives better insight into the generalizability of algorithms and how they would perform in a new clinical context [ 54 ]. However, external validation should not be considered the final step. A considerable number of authors have voiced the opinion that algorithms should be validated prospectively [ 55 ], perhaps even in a randomized clinical trial setting [ 7 , 52 , 56 ]. The number of randomized clinical trials for AI currently remains very low [ 57 , 58 ], but would allow algorithms to be validated on clinically meaningful endpoints [ 7 ], such as overall survival, improved quality of life or disease control for patients. Also, case read time reduction and user satisfaction for pathologists could be investigated, instead of the more computer vision-focused endpoints (such as pixel-wise accuracy, Dice score, specificity, or sensitivity) that AI papers tend to report.

The successful implementation of AI relies also on expectations and the experienced trustworthiness [ 59 ]: will the pathologist believe the presented data, especially if there is no visual control? Trustworthiness can be boosted by explainability, allowing a developer or user of the AI application to check which data to what extent contributed to the decision-making. In our example of predicting MYC translocations based on H&E-derived features [ 42 ], we could visualize through a technique that creates “saliency maps” [ 60 ] that the algorithm focuses on the areas involved by the diffuse large B-cell lymphoma. But this does not explain which characteristics of the involved tissue are important for the prediction. Explainable AI is a rather new field that potentially will not only help clinical acceptance, but is also important to prevent bias and will support discovery in pathology [ 61 ].

A different challenge is the flexibility of AI applications in pathology: will they be able to keep up with the needs of the field in an era of exploding molecular data and treatments? Self-learning algorithms that improve by doing [ 61 ] could be one answer. Availability of an array of task-specific algorithms which are integrated in the slide viewing software and can be chosen by the pathologist for specific tasks within the routine workflow seems to be a more realistic option for the near future. This process could be optimized by the use of computerized clinical decision support systems [ 62 ]. Any future scenario, however, requires pathologists having a basic understanding of AI, its potentials, and limitations.

Lastly, AI should be available in a fair way. Neither the available scanner, nor the used scanning software should be an obstacle to using AI algorithms, and accessibility, especially for hospitals with limited resources, should be a basic condition.

Conclusion

Currently, we see the first steps toward AI-based diagnostic support of hematopathology, focused on disease-specific tasks. Future applications are expected to integrate morphology with clinical and laboratory data and (cyto)genetics for a (differential) diagnosis, guiding additional tests and providing prognostic and predictive information based on the diagnosis. This requires large-scale training and validation. Supervision by a pathologist with basic understanding of AI, and visual control where possible, will be an essential aspect for clinical acceptance. Pathologists will not soon be replaced, but will instead work together with AI for a result greater than the sum of their parts.

Conflict of Interest Statement

Geert Litjens receives research funding from the Dutch Research Council, the Dutch Cancer Association, Health∼Holland, the European Research Council, the European Union, and the Innovative Medicine Initiative. He is on the advisory board of Canon Healthcare IT and is a shareholder of Aiosyn BV. L.V.E. and K.M.H. have no conflicts of interest to declare.

Funding Sources

The authors have no funding sources to declare for this research.

Author Contributions

L.V.E. and K.M.H. performed the literature analysis. K.M.H. drafted the paper. L.V.E., G.L., and K.M.H participated in writing the manuscript, reviewed the manuscript, and agreed with its contents.

Funding Statement

The authors have no funding sources to declare for this research.

References

1. Kaul V, Enslin S, Gross SA. History of artificial intelligence in medicine. Gastrointest Endosc. 2020;92((4)):807–812. doi: 10.1016/j.gie.2020.06.040.

2. Samuel AL. Some studies in machine learning using the game of checkers. IBM J Res Dev. 1959;3:210–229.

3. LeCun Y, Bengio Y, Hinton G. Deep learning. Nature. 2015;521((7553)):436–444. doi: 10.1038/nature14539.

4. Litjens G, Kooi T, Bejnordi BE, Setio AAA, Ciompi F, Ghafoorian M, et al. A survey on deep learning in medical image analysis. Med Image Anal. 2017;42:60–88. doi: 10.1016/j.media.2017.07.005.

5. Srinidhi CL, Ciga O, Martel AL. Deep neural network models for computational histopathology a survey. Med Image Anal. 2021;67:101813. doi: 10.1016/j.media.2020.101813.

6. Al-Janabi S, Huisman A, Van Diest PJ. Digital pathology current status and future perspectives. Histopathology. 2012;61:1–9. doi: 10.1111/j.1365-2559.2011.03814.x.

7. Kann BH, Hosny A, Aerts HJWL. Artificial intelligence for clinical oncology. Cancer Cell. 2021;39((7)):916–927. doi: 10.1016/j.ccell.2021.04.002.

8. van Eekelen L, Pinckaers H, van den Brand M, Hebeda KM, Litjens G. Using deep learning for quantification of cellularity and cell lineages in bone marrow biopsies and comparison to normal age-related variation. Pathology (Phila) 2022;54((3)):318–327. doi: 10.1016/j.pathol.2021.07.011.

9. Campanella G, Hanna MG, Geneslaw L, Miraflor A, Werneck Krauss Silva V, Busam KJ, et al. Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nat Med. 2019;25((8)):1301–1309. doi: 10.1038/s41591-019-0508-1.

10. Lu MY, Williamson DFK, Chen TY, Chen RJ, Barbieri M, Mahmood F. Data-efficient and weakly supervised computational pathology on whole-slide images. Nat Biomed Eng. 2021;5((6)):555–570. doi: 10.1038/s41551-020-00682-w.

11. Marini N, Marchesin S, Otálora S, Wodzinski M, Caputo A, van Rijthoven M, et al. Unleashing the potential of digital pathology data by training computer-aided diagnosis models without human annotations. NPJ Digit Med. 2022;5((1)):102. doi: 10.1038/s41746-022-00635-4.

12. Haferlach T, Pohlkamp C, Heo I, Drescher R, Hänselmann S, Lörch T, et al. Automated peripheral blood cell differentiation using artificial intelligence - a study with more than 10 000 routine samples in a specialized leukemia laboratory. Blood. 2021;138((Suppl 1)):103.

13. Wang CW, Huang SC, Lee YC, Shen YJ, Meng SI, Gaol JL. Deep learning for bone marrow cell detection and classification on whole-slide images. Med Image Anal. 2022;75:102270. doi: 10.1016/j.media.2021.102270.

14. El Achi H, Khoury JD. Artificial intelligence and digital microscopy applications in diagnostic hematopathology. Cancers (Basel) 2020;12((4)):797. doi: 10.3390/cancers12040797.

15. Baranova K, Tran C, Plantinga P, Sangle N. Evaluation of an open-source machine-learning tool to quantify bone marrow plasma cells. J Clin Pathol. 2021;74((7)):462–468. doi: 10.1136/jclinpath-2021-207524.

16. Fu F, Guenther A, Sakhdari A, McKee TD, Xia D. Deep learning accurately quantifies plasma cell percentages on cd138-stained bone marrow samples. J Pathol Inform. 2022;13:100011. doi: 10.1016/j.jpi.2022.100011.

17. Liang J, Malherbe JAJ, Fuller KA, Mirzai B, George C, Carter TL, et al. Automated enumeration of lymphoid and plasma cells in bone marrow to establish normal reference ranges. J Clin Pathol. 2018;71((10)):916–925. doi: 10.1136/jclinpath-2018-205168.

18. Song TH, Sanchez V, EIDaly H, Rajpoot NM. Dual-Channel active contour model for megakaryocytic cell segmentation in bone marrow trephine histology images. IEEE Trans Biomed Eng. 2017;64((12)):2913–2923. doi: 10.1109/TBME.2017.2690863.

19. Song TH, Sanchez V, Ei Daly H, Rajpoot NM. Simultaneous cell detection and classification in bone marrow histology images. IEEE J Biomed Health Inform. 2019;23((4)):1469–1476. doi: 10.1109/JBHI.2018.2878945.

20. Sirinukunwattana K, Aberdeen A, Theissen H, Sousos N, Psaila B, Mead AJ, et al. Artificial intelligence-based morphological fingerprinting of megakaryocytes a new tool for assessing disease in MPN patients. Blood Adv. 2020;4((14)):3284–3294. doi: 10.1182/bloodadvances.2020002230.

21. Nielsen FS, Pedersen MJ, Olsen MV, Larsen MS, Røge R, Jørgensen AS. Automatic bone marrow cellularity estimation in H&E stained whole slide images. Cytometry A. 2019;95((10)):1066–1074. doi: 10.1002/cyto.a.23885.

22. Tratwal J, Bekri D, Boussema C, Sarkis R, Kunz N, Koliqi T, et al. MarrowQuant across aging and aplasia a digital pathology workflow for quantification of bone marrow compartments in histological sections. Front Endocrinol (Lausanne) 2020;11:480. doi: 10.3389/fendo.2020.00480.

23. Zhou S, Marklund H, Blaha O, Desai M, Martin B, Bingham D, et al. Deep learning assistance for the histopathologic diagnosis of Helicobacter pylori. Intell Based Med. 2020;1–2:100004.

24. Steiner DF, MacDonald R, Liu Y, Truszkowski P, Hipp JD, Gammage C, et al. Impact of deep learning assistance on the histopathologic review of lymph nodes for metastatic breast cancer. Am J Surg Pathol. 2018;42((12)):1636–1646. doi: 10.1097/PAS.0000000000001151.

25. Bulten W, Balkenhol M, Belinga JA, Brilhante A, Çakır A, Egevad L, et al. Artificial intelligence assistance significantly improves Gleason grading of prostate biopsies by pathologists. Mod Pathol. 2021;34((3)):660–671. doi: 10.1038/s41379-020-0640-y.

26. Lee N, Jeong S, Park MJ, Song W. Deep learning application of the discrimination of bone marrow aspiration cells in patients with myelodysplastic syndromes. Sci Rep. 2022;12((1)):18677. doi: 10.1038/s41598-022-21887-w.

27. Kimura K, Tabe Y, Ai T, Matsuzaki A, Nishibe K, Ebihara M, et al. A novel automated image analysis system using deep convolutional neural networks to diagnose MDS. Blood. 2019;134((Supplement_1)):4670.

28. Brück OE, Lallukka-Brück SE, Hohtari HR, Ianevski A, Ebeling FT, Kovanen PE, et al. Machine learning of bone marrow histopathology identifies genetic and clinical determinants in patients with MDS. Blood Cancer Discov. 2021;2((3)):238–249. doi: 10.1158/2643-3230.BCD-20-0162.

29. Khoury JD, Solary E, Abla O, Akkari Y, Alaggio R, Apperley JF, et al. The 5th edition of the world health organization classification of haematolymphoid tumours myeloid and histiocytic/dendritic neoplasms. Leukemia. 2022;36((7)):1703–1719. doi: 10.1038/s41375-022-01613-1.

30. Thiele J, Kvasnicka HM, Diehl V. Standardization of bone marrow features-does it work in hematopathology for histological discrimination of different disease patterns? Histol Histopathol. 2005;20((2)):633–644. doi: 10.14670/HH-20.633.

31. Buhr T, Hebeda K, Kaloutsi V, Porwit A, van der Walt J, Kreipe H. European Bone Marrow Working Group trial on reproducibility of World Health Organization criteria to discriminate essential thrombocythemia from prefibrotic primary myelofibrosis. Haematologica. 2012;97((3)):360–365. doi: 10.3324/haematol.2011.047811.

32. Theissen H, Chakraborty T, Malacrino S, Royston D, Rittscher J. Multi-scale graphical representation of cell environment. Annu Int Conf IEEE Eng Med Biol Soc. 2022;2022:3522–3525. doi: 10.1109/EMBC48229.2022.9871710.

33. Sabattini E, Pizzi M, Agostinelli C, Bertuzzi C, Sagramoso Sacchetti CA, Palandri F, et al. Progression in ph-chromosome-negative myeloproliferative neoplasms an overview on pathologic issues and molecular determinants. Cancers (Basel) 2021;13((21)):5531. doi: 10.3390/cancers13215531.

34. Kvasnicka HM, Beham-Schmid C, Bob R, Dirnhofer S, Hussein K, Kreipe H, et al. Problems and pitfalls in grading of bone marrow fibrosis collagen deposition and osteosclerosis-a consensus-based study. Histopathology. 2016;68((6)):905–915. doi: 10.1111/his.12871.

35. Teman CJ, Wilson AR, Perkins SL, Hickman K, Prchal JT, Salama ME. Quantification of fibrosis and osteosclerosis in myeloproliferative neoplasms a computer-assisted image study. Leuk Res. 2010;34((7)):871–876. doi: 10.1016/j.leukres.2010.01.005.

36. Ryou H, Sirinukunwattana K, Aberdeen A, Grindstaff G, Stolz BJ, Byrne H, et al. Continuous indexing of fibrosis (CIF) improving the assessment and classification of MPN patients. Leukemia. 2023;37((2)):348–358. doi: 10.1038/s41375-022-01773-0.

37. Irshaid L, Bleiberg J, Weinberger E, Garritano J, Shallis RM, Patsenker J, et al. Histopathologic and machine deep learning criteria to predict lymphoma transformation in bone marrow biopsies. Arch Pathol Lab Med. 2022;146((2)):182–193. doi: 10.5858/arpa.2020-0510-OA.

38. Park BG, Chi HS, Jang S, Park CJ, Kim DY, Lee JH, et al. Association of cup-like nuclei in blasts with FLT3 and NPM1 mutations in acute myeloid leukemia. Ann Hematol. 2013;92((4)):451–457. doi: 10.1007/s00277-012-1645-5.

39. Bennett JM, Pryor J, Laughlin TS, Rothberg PG, Burack WR. Is the association of “cup-like” nuclei with mutation of the NPM1 gene in acute myeloid leukemia clinically useful? Am J Clin Pathol. 2010;134((4)):648–652. doi: 10.1309/AJCPULO8SLW0RKJL.

40. Eckardt JN, Middeke JM, Riechert S, Schmitt­mann T, Sulaiman AS, Kramer M, et al. Deep learning detects acute myeloid leukemia and predicts NPM1 mutation status from bone marrow smears. Leukemia. 2022;36((1)):111–118. doi: 10.1038/s41375-021-01408-w.

41. Eckardt JN, Schmittmann T, Riechert S, Kramer M, Sulaiman AS, Sockel K, et al. Deep learning identifies acute promyelocytic leukemia in bone marrow smears. BMC Cancer. 2022;22((1)):201. doi: 10.1186/s12885-022-09307-8.

42. Swiderska-Chadaj Z, Hebeda KM, van den Brand M, Litjens G. Artificial intelligence to detect MYC translocation in slides of diffuse large B-cell lymphoma. Virchows Arch. 2021;479((3)):617–621. doi: 10.1007/s00428-020-02931-4.

43. Wong WJ, Hasserjian RP, Pinkus GS, Breyfogle LJ, Mullally A, Pozdnyakova O. JAK2, MPL and ASXL1 mutational status correlates with distinct histological features in Philadelphia chromosome-negative myeloproliferative neoplasms. Haematologica. 2018;103((2)):e63–e68. doi: 10.3324/haematol.2017.178988.

44. van Dam S, Baars MJD, Vercoulen Y. Multiplex tissue imaging spatial revelations in the tumor microenvironment. Cancers (Basel) 2022;14((13)):3170. doi: 10.3390/cancers14133170.

45. Walters DK, Jelinek DF. Multiplex immunofluorescence of bone marrow core biopsies visualizing the bone marrow immune contexture. J Histochem Cytochem. 2020;68((2)):99–112. doi: 10.1369/0022155419896802.

46. Kristensen HB, Andersen TL, Patriarca A, Kallenbach K, MacDonald B, Sikjaer T, et al. Human hematopoietic microenvironments. PLoS One. 2021;16((4)):e0250081. doi: 10.1371/journal.pone.0250081.

47. Brück O, Dufva O, Hohtari H, Blom S, Turkki R, Ilander M, et al. Immune profiles in acute myeloid leukemia bone marrow associate with patient age T-cell receptor clonality and survival. Blood Adv. 2020;4((2)):274–286. doi: 10.1182/bloodadvances.2019000792.

48. Patel SS, Lipschitz M, Pinkus GS, Weirather JL, Pozdnyakova O, Mason EF, et al. Multiparametric in situ imaging of NPM1-mutated acute myeloid leukemia reveals prognostically-relevant features of the marrow microenvironment. Mod Pathol. 2020;33((7)):1380–1388. doi: 10.1038/s41379-020-0498-z.

49. Ninkovic S, Purton LE, Harrison SJ, Quach H. Assessing the immune tumour microenvironment (iTME) using multiplex immunoflourescence histochemistry (mIHC) demonstrates close proximity of cytotoxic T-cells to plasma cells (PC) in patients with newly diagnosed multiple myeloma (NDMM) Blood. 2021;138((Suppl 1)):4705.

50. Linmans J, Elfwing S, van der Laak J, Litjens G. Predictive uncertainty estimation for out-of-distribution detection in digital pathology. Med Image Anal. 2023;83:102655. doi: 10.1016/j.media.2022.102655.

51. Walter W, Haferlach C, Nadarajah N, Schmidts I, Kühn C, Kern W, et al. How artificial intelligence might disrupt diagnostics in hematology in the near future. Oncogene. 2021;40((25)):4271–4280. doi: 10.1038/s41388-021-01861-y.

52. van der Laak J, Litjens G, Ciompi F. Deep learning in histopathology the path to the clinic. Nat Med. 2021;27((5)):775–784. doi: 10.1038/s41591-021-01343-4.

53. Tellez D, Litjens G, Bándi P, Bulten W, Bokhorst JM, Ciompi F, et al. Quantifying the effects of data augmentation and stain color normalization in convolutional neural networks for computational pathology. Med Image Anal. 2019;58:101544. doi: 10.1016/j.media.2019.101544.

54. Staartjes VE, Kernbach JM. Significance of external validation in clinical machine learning let loose too early? Spine J. 2020;20((7)):1159–1160. doi: 10.1016/j.spinee.2020.02.016.

55. Kang J, Morin O, Hong JC. Closing the gap between machine learning and clinical cancer care-first steps into a larger world. JAMA Oncol. 2020;6((11)):1731–1732. doi: 10.1001/jamaoncol.2020.4314.

56. Kelly CJ, Karthikesalingam A, Suleyman M, Corrado G, King D. Key challenges for delivering clinical impact with artificial intelligence. BMC Med. 2019;17((1)):195. doi: 10.1186/s12916-019-1426-2.

57. Nagendran M, Chen Y, Lovejoy CA, Gordon AC, Komorowski M, Harvey H, et al. Artificial intelligence versus clinicians systematic review of design, reporting standards, and claims of deep learning studies. BMJ. 2020;368:m689. doi: 10.1136/bmj.m689.

58. Dong J, Geng Y, Lu D, Li B, Tian L, Lin D, et al. Clinical trials for artificial intelligence in cancer diagnosis a cross-sectional study of registered trials in ClinicalTrials.gov. Front Oncol. 2020;10:1629. doi: 10.3389/fonc.2020.01629.

59. Drogt J, Milota M, Vos S, Bredenoord A, Jongsma K. Integrating artificial intelligence in pathology a qualitative interview study of users' experiences and expectations. Mod Pathol. 2022;35((11)):1540–1550. doi: 10.1038/s41379-022-01123-6.

60. Simonyan K, Vedaldi A, Zisserman A. Deep inside convolutional networks visualising image classification models and saliency maps. InProceedings of the international conference on learning representations (ICLR) ICLR. 2014 apr 16;:2014.

61. Krajna A, Kovac M, Brcic M, Šarčević A. Explainable artificial intelligence an updated perspective. In: 2022 45th jubilee. International convention on information communication and electronic technology MIPRO. 2022:p. 859–864.

62. Sutton RT, Pincock D, Baumgart DC, Sadowski DC, Fedorak RN, Kroeker KI. An overview of clinical decision support systems benefits, risks, and strategies for success. NPJ Digit Med. 2020;3:17. doi: 10.1038/s41746-020-0221-y.