Digital pathology (DP) built on whole-slide imaging (WSI) technology has fundamentally changed how pathologists analyze tissue specimens. Modern scanners produce 20x to 40x digital slices at extremely fast scan rates, generating high-resolution images with rich histological detail. For hematopathology specifically, the demands are even higher: peripheral blood and bone marrow (BM) smears require 60x or 100x magnification scans to extract reliable information on cell morphology and pathological features. This volume of high-resolution image data creates both the need and the opportunity for AI-assisted analysis.
What AI brings to the table: Artificial intelligence, as a branch of computer science, attempts to replicate human intelligent behavior through machines. Two factors have historically limited AI development: the lack of sizable labeled datasets and insufficient computing power. Both constraints have eased substantially in recent years. Advances in image acquisition technology have created massive repositories of pathology images, while continuous improvements in GPU computing power have made large-scale model training practical. Commercial AI tools developed for solid tumor anatomical pathology now include platforms such as Lunit, Ibex, aetherAI, DeepBio, PathAI, and Paige.AI.
Hematology-specific tools: AI development in blood diseases has followed a parallel track. DeepFlow, developed by DeepAnalysis Intelligence, is described as the world's first flow cytometry AI cloud diagnosis system. Its accuracy rate for acute leukemia diagnosis reaches 95%, and it operates approximately 100 times faster than manual physician review. This kind of performance illustrates the potential impact of AI across the hematology spectrum: not only matching human expert accuracy but doing so at a speed that is operationally transformative for high-volume laboratories.
This 2022 review from Blood Science surveys AI applications across the hematopathology pipeline, covering normal hematopoietic cell identification, heterogeneous blood cell classification, acute leukemia morphological recognition, and clonal hematopoietic stem cell disorder diagnosis. It also provides guidance for researchers designing and evaluating machine learning studies in hematology, covering model architecture choices, dataset construction, and performance evaluation standards.
Machine learning is broadly defined as the ability of machines to mimic human intelligent behavior by extracting meaningful results from complex data through computation. Three subcategories exist: supervised learning (training on labeled examples), unsupervised learning (finding structure in unlabeled data), and reinforcement learning (learning through feedback). Within supervised learning, common methods include regression, support vector machines (SVM), decision trees, and neural networks. In hematopathology, supervised methods dominate because expert-annotated cell images provide the labeled ground truth needed for training classifiers.
Deep learning: Deep learning is a subfield of machine learning that emphasizes multi-layer neural network architectures designed to simulate how the human brain processes information. Deep neural networks are composed of multilayer interconnected nodes, where each layer builds on the previous to improve and optimize prediction or classification. The computational process of passing data forward through these layers is called forward propagation. The input layer handles data preprocessing, and the output layer produces the final prediction or classification.
Key architectures: Convolutional neural networks (CNNs) are the dominant architecture for image analysis tasks in hematopathology. CNNs learn hierarchical feature representations directly from pixel data, detecting edges and textures in early layers and increasingly abstract features (cell shapes, nuclear chromatin patterns, cytoplasmic characteristics) in deeper layers. Recurrent neural networks (RNNs) and Long Short-Term Memory (LSTM) networks are used for sequential data, such as tracking cell behavior over time in time-lapse microscopy. In hematopathology applications, CNN and RNN architectures are often combined to capture both spatial image features and temporal cell dynamics simultaneously.
The expanding availability of digitized WSI data has been critical to training these models effectively. Large annotated datasets provide the volume of labeled examples needed to train deep learning models that generalize across patient populations, staining protocols, and imaging hardware. The improvement in GPU computing power has made it feasible to process the large image files and run multiple images in parallel, substantially reducing training and inference time for clinical deployment.
Accurate segmentation and identification of blood cells is the foundational step for extracting features used in diagnosing both benign and malignant hematologic diseases. Manual counting of hematopoietic cells in microscopic images is tedious, time-consuming, and highly subjective, making it a natural target for automation. Early work in this area focused on white blood cell (WBC) and red blood cell (RBC) segmentation. In 2014, Alomari et al. proposed an iterative structured circle detection algorithm for automated WBC and RBC counting, achieving average accuracies of 95.3% for RBCs and 98.4% for WBCs using a thresholding-based image separation approach.
Deep learning for WBC identification: The first application of deep learning to WBC identification came from Shahin et al. in 2017, who developed a CNN-based identification system for WBCs in blood smear images achieving an overall accuracy of 96.1%. Building on this, Wang et al. applied two object detection architectures - Single Shot Multibox Detector (SSD) and an incremental improvement of You Only Look Once (YOLO) - to leukocyte identification. Their model was trained on 14,700 annotated images and tested on 1,120 annotated images covering 11 categories of peripheral leukocytes, achieving a best mean average precision of 93.10% and mean accuracy of 90.09%. In 2019, Hegde et al. made the first direct comparison between traditional image processing and deep learning approaches for WBC classification, with CNN achieving the highest accuracy of 99% against traditional methods.
Domain adaptation: A persistent problem in WBC classification models is domain shift - severe performance degradation when models trained on one institution's data are tested on data from a different setting with different staining protocols or scanner hardware. Pandey et al. addressed this using unsupervised domain adaptation (UDA) techniques, where a "nearest clone" from source data acts as a proxy for classifying target domain images. Lippeveld et al. took a different approach, using stain-free imaging flow cytometry to classify WBC types, comparing two deep learning and two classical machine learning methods - and notably finding that deep learning approaches did not outperform methods based on manually engineered features in the stain-free imaging context.
Predicting cell fate: Buggenthin et al. pushed beyond current cell-state classification to predicting future cell behavior, developing a deep neural network that prospectively predicts hematopoietic lineage choice using time-lapse brightfield microscopy and cell movement data. By combining CNN and RNN architectures to correlate cell morphology from brightfield images with lineage selection data from fluorescence microscopy, their method identified stem cell differentiation trajectories more efficiently than traditional approaches while eliminating the need for fluorescently labeled samples.
Platelets are produced by megakaryocytes in the bone marrow and play a central role in hemostasis and thrombotic disease. Deficiencies in platelet function produce a large and varied spectrum of bleeding disorders ranging from manageable skin and mucosal bruising to organ tissue damage and fatality. Identifying abnormal megakaryocyte morphology is particularly important for diagnosing Philadelphia-negative myeloproliferative neoplasms (MPNs), which include essential thrombocythemia (ET), polycythemia vera (PV), and primary myelofibrosis (PMF). These three conditions share overlapping clinical and laboratory features, making distinction challenging at early disease time points, and the cytological and topographic features of megakaryocytes in bone marrow trephines (BMTs) are central to histological interpretation.
Automated megakaryocyte analysis: Sirinukunwattana et al. developed a multi-step machine learning pipeline for automated identification, quantitative analysis, and abstract representation of megakaryocyte features on digital images of hematoxylin and eosin (H&E)-stained bone marrow sections. The pipeline uses three sequential components: a Single Shot Multibox Detector deep neural network to predict megakaryocyte locations in the slide; U-Net segmentation to partition image regions containing megakaryocyte cells; and an autoencoder neural network to learn a numerical vector encoding of megakaryocyte cytomorphology - a compact feature representation that captures cell shape, size, nuclear lobulation, and cytoplasmic characteristics. This method achieved a high predictive accuracy with an area under the curve (AUC) of 0.95 for distinguishing between MPN disease states.
Sickle cell disease and erythrocyte shape classification: Many blood diseases involve significant changes in RBC shape and size, and these morphological changes reflect protein-level alterations that affect cell function. Sickle cell disease (SCD) is a prominent example: in addition to the characteristic sickle shape, RBCs in SCD display numerous shape variants directly related to the degree of sickle hemoglobin polymerization. Xu et al. developed a deep convolutional network framework to classify the heterogeneous RBC shapes present in sickle blood, achieving a high recall of 93.8% but a relatively modest precision of 85.0%. This precision-recall asymmetry - high sensitivity with more false positives - reflects the inherent difficulty of distinguishing the most ambiguous intermediate cell shapes.
Together, these studies validate the feasibility of applying machine learning to objective, quantitative descriptions of blood cell morphology in disorders where subjective human assessment has historically limited diagnostic reproducibility and standardization across laboratories.
Leukemia originates in the bone marrow through the overproduction of immature leukocytes that fail to mature and crowd out normal blood cells. Acute myeloid leukemia (AML) is among the most common subtypes of acute leukemia, defined by the accumulation of myeloid blasts in the BM. Morphological identification of blast cells has historically been a powerful tool for hematologists, but it is subjective and demanding even for expert reviewers. The first step in AML diagnosis is careful microscopic examination of stained blood smears or BM aspirates, making automated image analysis a direct target for improvement.
Early SVM-based approaches: Kazemi et al. developed one of the first computational frameworks for AML classification and subtype discrimination using support vector machines (SVM). Their automated pipeline performs image preprocessing, cell segmentation, post-processing, feature extraction, and final SVM-based classification, achieving sensitivity of 95%, specificity of 98%, and overall accuracy of 96% on blood microscopic images. This established the performance baseline that subsequent deep learning approaches would seek to exceed.
Deep learning at human expert level: Matek et al. assembled an annotated dataset of over 18,000 WBCs from peripheral blood smears and trained a CNN for leukocyte classification. When evaluated against inter- and intra-expert variability, the network attained precision and sensitivity above 90% for the most common physiological leukocyte classes as well as for myeloblasts, the diagnostically critical blast cell population in AML. This demonstrated human-level performance in blast cell recognition. In a subsequent study, Matek et al. compiled the largest expert-annotated BM cytology image dataset in the literature at the time, containing 171,374 microscopic cytological images from 945 patients diagnosed with a range of hematological diseases. CNN classifiers trained on this dataset identified diagnostically relevant cell species with high precision and recall, outperforming previous feature-based methods.
Multi-myeloma and imaging modalities: Beyond peripheral blood and BM smear images, AI has been applied to CT and PET/CT imaging for myeloid malignancies. Martinez-Martinez et al. developed a fully automated classification system for bone marrow infiltration in low-dose CT of multiple myeloma (MM) patients using a probabilistic density model with supervised learning. Xu et al. applied deep learning to automated whole-body bone lesion detection for MM using Ga-68-pentixafor PET/CT, with a V-Net 3D segmentation architecture achieving 99.51% specificity and AUC of 0.99 for lesion detection.
AI classification of lymphoid malignancies presents distinct challenges compared to AML because the morphological changes in lymphocytic cells are often subtler. For acute lymphoblastic leukemia (ALL), the blasts can be more difficult to distinguish from reactive lymphoid populations than AML blasts from normal myeloid precursors. The field has accordingly moved toward increasingly sophisticated deep learning architectures to capture these subtle morphological differences.
Pretrained network transfer learning for ALL: Shafique et al. demonstrated the power of transfer learning for ALL detection by deploying the pretrained AlexNet CNN, fine-tuned on their own training dataset. This approach avoids training a network from scratch and leverages representations learned from large natural image datasets. The framework achieved a sensitivity of 100%, specificity of 98.11%, and accuracy of 99.50% for ALL detection. For ALL subtype classification - a more challenging multi-class problem - it achieved sensitivity of 96.74%, specificity of 99.03%, and accuracy of 96.06%. Rehman et al. followed with a comparison of AlexNet, Naive Bayesian, K-Nearest Neighbor, and SVM for ALL subtype classification, with AlexNet achieving the best accuracy of 97.78%.
Lymphoma classification: For non-Hodgkin lymphoma subtypes, El Achi et al. developed a CNN for automated lymphoma diagnosis from digital pathology images, achieving 95% accuracy in distinguishing lymphoma from non-lymphoma and across lymphoma subtypes. Mohlman et al. focused specifically on the clinically important distinction between Burkitt lymphoma (BL) and diffuse large B-cell lymphoma (DLBCL), two high-grade lymphomas with very different treatment implications. Their AI system to assist human decision-making (augmented intelligence) achieved 97.1% accuracy on microscopic images. Syrykh et al. applied Bayesian neural network approaches to WSI-based lymphoma classification across a broader set of subtypes, achieving AUC of 0.99 with a training set of 160,000 images.
Performance trends across Table 2: Across all the lymphoid malignancy studies summarized in the review, published between 2014 and 2020, performance metrics consistently cluster in the 93-98% accuracy range for binary classification tasks and 85-97% for multi-subtype problems. Earlier studies relied on SVM, K-Nearest Neighbor (K-NN), and Naive Bayes classifiers applied to manually engineered features. Later studies transitioned to CNNs (VGGNet, AlexNet, custom architectures) and achieved uniformly higher performance, with the most recent Bayesian neural network approach reaching AUC 0.99 on the largest dataset.
Clonal hematopoietic stem cell disorders include aplastic anemia, myeloproliferative diseases, myelodysplastic syndromes (MDS), and paroxysmal nocturnal hemoglobinuria. MDS is a hematopoietic tumor characterized by abnormal, ineffective hematopoiesis and its diagnosis relies primarily on morphological manifestations - making it a strong candidate for AI-assisted image analysis. MDS diagnosis requires expert assessment of multiple peripheral blood and bone marrow features simultaneously, including dysplastic changes across multiple cell lineages, and the complexity of this assessment makes standardization across centers particularly challenging.
The Kimura et al. automated MDS diagnostic system: Kimura et al. developed an automated diagnostic support system for MDS by combining two AI components: a deep CNN-based automated blood cell image recognition system and a decision-making system using extreme gradient boosting (XGBoost). The CNN handles cell-level image classification, while XGBoost integrates the cell classification outputs with clinical context to produce a final diagnostic decision. The model was trained on a dataset of 695,030 blood cell images extracted from 3,261 peripheral blood smears, spanning a broad range of hematological malignancies. This scale - nearly 700,000 individual cell images - represents one of the largest training datasets described in the review.
Multi-class morphological classification: The trained model can simultaneously classify 17 distinct blood cell types and 97 separate morphological features of those cells. Performance was validated at greater than 93.5% sensitivity and greater than 96.0% specificity across these 17 cell type categories. This multi-dimensional classification capacity is clinically significant because MDS diagnosis requires recognizing dysplastic changes across erythroid, myeloid, and megakaryocytic lineages simultaneously, not just identifying a single abnormal cell type. The ability to detect morphological abnormalities across 97 features in a single automated pass represents a qualitative advance over systems that target only one or two cell characteristics.
This approach establishes a proof of concept for AI-assisted MDS diagnosis at scale, potentially enabling more consistent morphological assessment across institutions with varying levels of hematopathology expertise. The integration of CNN-based cell recognition with XGBoost decision-making illustrates a general design pattern: deep learning for feature extraction from raw images combined with gradient-boosted trees or other interpretable classifiers for final decision-making, providing a degree of model transparency not available with pure end-to-end deep learning.
The labeled data bottleneck: Despite strong performance metrics, the authors emphasize that insufficient labeled training data remains a major limiting factor for AI in hematopathology. Generating expert-annotated cell images requires the time and expertise of trained hematopathologists, making large-scale annotation expensive and slow. The cost of expert annotation and the sheer number of images required to train robust deep learning models are identified as the primary factors constraining model training quality, particularly for rare cell types and uncommon disease entities where fewer labeled examples exist.
Model opacity: Deep learning models require large numbers of parameters - including network topology, initial weight values, and thresholds - and their internal learning process cannot be directly observed. The output results are correspondingly difficult to explain to clinicians, which affects the reliability and acceptability of AI-generated diagnoses. This "black box" problem is a recognized obstacle to clinical adoption: pathologists and hematologists need to understand why a model reached a particular classification, not just that it did so with high accuracy. Explainability methods such as saliency maps and gradient-weighted class activation mapping (Grad-CAM) are available but provide only partial transparency into CNN decision-making.
Generalization and standardization gaps: Most published AI hematopathology models are developed and validated at single institutions using datasets from a specific scanner, staining protocol, and patient population. Performance on externally acquired data from different institutions with different equipment consistently drops relative to internal validation. The lack of standardized reporting frameworks for AI studies in hematopathology makes it difficult to compare results across papers and assess the actual clinical readiness of any given tool.
Future directions: The authors see deep learning as having strong potential to push beyond the limits of traditional microscopy by enhancing resolution, signal content, and information extraction from acquired image data. Expanding the use of WSI across hematopathology practice will create larger and more diverse training datasets. Integration of AI analysis with multimodal clinical data - combining morphological image features with genomic, flow cytometric, and clinical variables - is identified as the most promising direction for improving diagnostic precision and clinical utility. Federated learning frameworks enabling multi-institutional model training without sharing raw patient data could address the generalization problem while respecting data privacy constraints.