Deep learning (DL) is a subset of machine learning that uses multilayered neural networks to automatically extract subtle graphical features from images and produce highly accurate predictions. Over the past decade it has reshaped tasks in computer vision, facial recognition, and object detection, and researchers have begun applying it systematically to medical imaging, including radiology scans and digitized pathology specimens. This 2023 systematic review and visual glossary, published in Haematologica, is one of the most comprehensive surveys of DL applications in hematology to date, covering both malignant conditions (lymphomas, leukemias, myeloma, myelodysplastic syndrome) and non-malignant conditions (anemia, malaria, sickle cell disease, pulmonary embolism).
Why hematology specifically: Visual assessment of digitized biospecimens and clinical imaging is central to every phase of hematologic care, from peripheral blood smear review and bone marrow aspirate interpretation to PET/CT staging and treatment response assessment. Because hematologic diseases produce characteristic morphological changes in blood cells, marrow architecture, and lymph node structure, image-based DL has natural clinical relevance. Despite this, DL applications in hematology lag behind other oncology domains, largely because of limited publicly annotated multi-modal datasets and the rarity of individual disease entities.
Scope and structure: The authors conducted a PubMed Boolean query spanning January 1990 through August 2022, combining "deep learning" terms with "hematology" terms, yielding 2,708 initial articles. After manual abstract review and full-text assessment, 65 manuscripts were included in the final analysis. These manuscripts are organized by clinical care phase, covering task automation, detail optimization, disease detection, differential diagnosis, disease classification, risk prediction, complication assessment, therapy response, and survival and relapse prediction. The paper also provides a visual glossary explaining core DL concepts, including neural network structure, convolution operations, transfer learning, multiple-instance learning, Vision Transformers, and explainability methods such as Grad-CAM, UMAP, and t-SNE.
The review targets practicing hematologists and hematopathologists who need to critically evaluate DL literature without a computer science background. The visual glossary and structured clinical framework make it a particularly accessible entry point into an otherwise technically dense field.
A neural network processes input images by converting pixel values into numerical arrays and propagating them through successive layers of nodes. Each node applies a mathematical transformation to represent progressively more complex image features: earlier layers detect edges and textures, while deeper layers encode higher-level structures such as cellular shapes, nuclear contours, or tissue architecture. The connections between nodes are governed by trainable parameters that are iteratively adjusted during training through forward propagation (image to prediction) and back propagation (adjusting weights based on prediction error using a loss function and optimizer algorithm). Given that a single forward pass may require over 10^9 arithmetic operations, training typically demands Graphics Processing Unit (GPU) hardware.
Convolutional neural networks (CNNs): The dominant architecture for image-based DL applies the convolution operation between layers, multiplying overlapping sub-sections of the input image by trainable filter matrices to produce lower-dimensional feature representations. Residual Neural Networks (ResNets), which use "skip connections" allowing information to bypass certain layers, are the most commonly used CNN class in hematology research. Specific variants including Inception, EfficientNets, and MobileNets are open-source and widely available. Compared to shallow networks, ResNets can be trained to greater depth without the vanishing gradient problem.
Transfer learning: Training a CNN from randomly initialized parameters requires millions of labeled images and weeks of computation on industry-grade hardware. In clinical research, these resources are rarely available. Transfer learning addresses this by starting with a network pre-trained on a large general-purpose dataset (such as ImageNet, containing over 1 million images) whose early layers have already learned general low-level features. Only the final layers are retrained on the clinical task, dramatically reducing the required dataset size and computational cost. This is the dominant training strategy across all 65 reviewed manuscripts.
Multiple Instance Learning (MIL) and Vision Transformers (ViT): MIL frameworks address the challenge of weak supervision, where a single label applies to an entire whole-slide image rather than individual cells or patches. The image is divided into smaller subsets (bags), and the model predicts "positive" if at least one patch is positive. Attention-based MIL variants, including the Clustering-constrained Attention MIL (CLAM) framework, assign numeric weights to patches to produce explainable attention heatmaps. Vision Transformers, in contrast, avoid convolution entirely and instead analyze a grid of image patches in parallel, capturing global spatial relationships. ViTs match CNNs at current clinical dataset scales and outperform them at larger scales while requiring fewer parameters.
The systematic search combined one "deep learning" term (such as "deep learning," "convolutional neural network," or "neural network") with one "hematology" term (such as "leukemia," "lymphoma," "myeloma," "anemia," or "myelodysplastic syndrome") in PubMed across all publications from January 1, 1990 through August 1, 2022. The initial query returned 2,708 articles. One author performed manual review of titles and abstracts to exclude non-clinical applications, review articles without original data, non-image-based DL work, and publications outside the scope of hematologic conditions, reducing the pool to 237 eligible manuscripts. Full-text review of these 237 manuscripts yielded a final 65 articles for inclusion.
Exclusion categories: Articles were excluded if they used non-image-based neural networks (e.g., models operating solely on tabular clinical data), if they applied machine learning architectures without neural network components, or if their primary application was outside hematology (e.g., using hematologic data as a control group for another disease). This structured approach, following a PRISMA-style diagram, ensured that included manuscripts all involved image-based DL applied to a recognized hematologic condition in a clinically relevant context.
Performance metrics used across studies: Classification models are most frequently evaluated by the Area Under the Receiver Operating Characteristic Curve (AUROC), with values near 1.0 indicating near-perfect discriminatory power and values near 0.5 indicating no better than chance. Sensitivity and specificity are reported alongside AUROC when threshold-specific performance is clinically meaningful. For segmentation tasks, the Sorensen-Dice similarity coefficient (Dice) measures the spatial overlap between predicted and ground truth regions of interest, where a Dice of 1.0 represents perfect overlap. The Jaccard index (Intersection over Union, IoU) is a related segmentation metric also used in several studies. AUROC was the most common primary outcome metric, appearing in the large majority of the 65 reviewed papers.
Critical appraisal framework: The authors evaluated each manuscript across four dimensions: primary performance metric and value, validation strategy (internal cross-validation vs. external test cohort vs. prospective validation), explainability method (Grad-CAM, saliency maps, UMAP, t-SNE, or none), and presence of direct comparison against human expert performance. This framework, tabulated across four comprehensive supplementary tables, reveals clear patterns in where the field is strong and where critical gaps remain before clinical translation.
A substantial portion of the reviewed work addresses the automation of repetitive, time-consuming clinical tasks that currently require significant human labor. In hematopathology workflows, DL models trained to contour white blood cell (WBC) borders in peripheral blood smears achieved near-perfect Dice coefficients across multiple cohorts, enabling downstream automated cell counts with high accuracy. Similarly, DL automates the identification and rotation of chromosomes in digital karyograms, a process that is otherwise manually intensive for every hematologic malignancy workup requiring cytogenetic analysis.
Lesion localization on radiology: In radiology workflows, DL models have been applied to localize and segment hypermetabolic lesions on PET/CT for several lymphoma subtypes and for multiple myeloma. Models localizing FDG-avid lesions in diffuse large B-cell lymphoma (DLBCL) and pediatric Hodgkin lymphoma on PET/CT achieved Dice coefficients ranging from 0.86 to 0.98 across the reviewed studies. Automated spleen volume calculation from CT scans in myeloproliferative neoplasm (MPN) patients and automated clot burden quantification from CT pulmonary angiography (CTPA) in pulmonary embolism were also demonstrated, with the latter having direct potential to reduce radiologist workload in time-sensitive emergency workflows.
Detail optimization: DL-based image enhancement has been explored to either improve diagnostic readability or reduce patient exposure to toxic agents. For sickle cell disease, U-Net architectures upscaled mobile-device photos of peripheral blood smears to match laboratory microscope quality, retaining relevant visual cues with near-perfect downstream classification performance. For lymphoma patients undergoing PET-MRI, DL augmentation demonstrated the feasibility of using reduced 18F-FDG doses while maintaining diagnostic quality, and for multiple myeloma on CT, DL-enhanced low-dose images maintained diagnostic quality at proposed dose reductions. These optimizations carry both clinical and safety benefits by reducing nephrotoxic contrast exposure and radioactive tracer doses.
Notably, attempts to enhance peripheral blood images for malaria RBC inclusion detection were less successful: CNN-based upscaling was insufficient to resolve parasites that were not already clearly visible at lower resolution, illustrating that enhancement models have modality-specific limits and that performance on one imaging task does not necessarily transfer to morphologically different targets.
Disease detection represents the largest single category of reviewed DL applications. In peripheral blood smear analysis, DL models have demonstrated high accuracy for detecting Plasmodium parasites associated with malaria across both cross-validated internal cohorts and external validation datasets, with AUROC values reaching 0.95-0.99 in multiple studies. Hemoglobin H inclusions in alpha-thalassemia are detectable from appropriately stained peripheral blood with DL methods. For transfusion medicine, DL applied to stored RBC images can objectively assess quality and degradation prior to transfusion, with explainability analyses identifying novel morphological features associated with poor storage outcomes that had not been previously characterized.
WBC abnormality detection: Dysplastic neutrophils pathognomonic for myelodysplastic syndrome (MDS) are consistently detected by DL from peripheral blood smears with high accuracy. Models trained on bone marrow aspirate images detect cellular morphologies associated with MPN, acute promyelocytic leukemia (APL), and multiple myeloma. For APL specifically, rapid accurate recognition is clinically urgent because the disease requires emergent treatment with all-trans retinoic acid and is associated with life-threatening coagulopathy if diagnosis is delayed. DL models detecting APL-specific promyelocyte morphologies from bone marrow smears and peripheral blood smears achieved AUROCs of 0.97-0.99 in several reviewed studies.
Lymphoma detection on radiology: DL applied to PET/CT images achieves high classification performance for DLBCL hypermetabolic lesions. Models detecting mantle cell lymphoma (MCL) from PET/CT face more challenging tradeoffs between sensitivity and false positive rates in external cohorts, highlighting generalizability concerns. For non-malignant hematologic emergencies, DL models demonstrate strong performance for pulmonary embolism (PE) detection on CT pulmonary angiography, with one multimodal model integrating CT images and electronic health record (EHR) data outperforming image-only models. Cost analyses in DVT and PE studies demonstrated positive financial benefit to healthcare systems from automated detection, adding an economic rationale to the clinical case.
Atypical imaging modalities: Several studies demonstrated that anemia can be detected with high accuracy from non-traditional imaging modalities. DL applied to electrocardiograms (ECG) detected anemia with an AUROC of approximately 0.88, exploiting subtle QRS complex changes caused by compensatory cardiac remodeling. Separately, DL on funduscopic examination images detected anemia from retinal features such as optic disk pallor. Both approaches employ explainability methods that reveal the specific image features driving predictions, offering plausible biological mechanisms and validating that models are not relying on artifacts.
Differential diagnosis in hematology is complicated by diseases that share overlapping morphological presentations requiring integration of clinical context, immunohistochemistry, flow cytometry, and molecular data for definitive characterization. DL models have been developed to assist with several of these classification challenges using image data alone. One striking example involves transfer learning pre-trained on images of common household objects (bananas, rings, pears) to identify analogously shaped RBC inclusions of Toxoplasma, Plasmodium, and Babesia, leveraging the morphological similarity between biological inclusions and everyday objects to bootstrap performance with limited labeled clinical data.
Leukemia differential diagnosis: Distinguishing acute myeloid leukemia (AML) from acute lymphoblastic leukemia (ALL) typically requires flow cytometry for immunophenotyping, since visual blast morphology alone is considered insufficient. Multiple DL models have demonstrated the ability to differentiate AML from ALL using only peripheral blood or bone marrow aspirate images, achieving AUROCs above 0.90 in internally validated cohorts. DL has also been applied to classify MYC gene rearrangements in DLBCL patients from routine H&E lymph node biopsy images, a task typically requiring fluorescent in situ hybridization (FISH). The model maintained high accuracy on external cohorts, suggesting it is learning biologically relevant morphological correlates of this genomic alteration rather than site-specific artifacts.
Lymphoma subtyping: Distinguishing between NHL subtypes, including DLBCL, Burkitt lymphoma, follicular lymphoma (FL), and small lymphocytic lymphoma, from standard H&E lymph node biopsies is a challenging differential with direct treatment implications. DL models differentiating among these subtypes have achieved AUROCs of 0.84-0.99 in reviewed studies. Li et al. trained a DLBCL diagnostic model that maintained high accuracy across four separate institutional cohorts, demonstrating reproducibility that is rare in the DL hematology literature. Syrykh et al. developed a novel uncertainty quantification approach for FL versus reactive follicular hyperplasia, reporting higher classification accuracy when restricting predictions to the low-uncertainty subset of images.
WBC and RBC classification: DL models classifying WBC morphologies in routine blood smears consistently achieve accuracy above 90% with explainability techniques highlighting sensitive cellular features such as nuclear lobulation, granularity, and chromatin texture. Sickle cell disease RBC classification and platelet agonist identity prediction from imaging flow cytometry have also been demonstrated with high accuracy, illustrating the breadth of classification tasks amenable to DL in hematology. However, the majority of WBC classification studies report only internal validation, and the few with external validation show measurable performance drops.
Beyond diagnosis and classification, a smaller but growing body of work applies DL to later phases of clinical care where the clinical stakes are high and the predictive landscape is correspondingly complex. Risk prediction has been explored using CT imaging and digitized bone marrow biopsies for DLBCL outcomes. DL models have demonstrated the ability to predict transformation of low-grade lymphomas (chronic lymphocytic leukemia and follicular lymphoma) to high-grade DLBCL using bone marrow biopsy images, with AUROC around 0.73-0.86 internally. Separately, CT-based segmentation and quantification of muscle tissue as a proxy for sarcopenia has been used as a prognostic marker in DLBCL patients, with DL-quantified sarcopenia independently predicting mortality in patients under 60 years.
Complication assessment: Cai et al. developed a DL model detecting sea fan neovascularization from ultra-widefield color fundus photographs in sickle cell disease patients, a vision-threatening complication that warrants prophylactic management. The model achieved an AUROC of 0.99 on internal validation, and its clinical utility lies in enabling screening during routine ophthalmology visits without requiring specialized retinal specialists to identify this specific sickle cell complication. For pulmonary embolism, a DL model classifying right ventricular strain from volumetric CTPA scans (AUROC 0.88 internally) provides quantitative severity assessment that goes beyond simple PE detection.
Therapy response and minimal residual disease: Doan et al. applied a label-free computer vision approach to detect residual ALL lymphoblasts after induction chemotherapy using imaging flow cytometry, achieving an accuracy of 0.88. This approach to minimal residual disease (MRD) assessment is significant because MRD status after induction is one of the strongest predictors of relapse in ALL, and label-free methods avoid the need for immunophenotypic staining that conventional flow cytometry requires.
Survival and relapse prediction: DL models predicting relapse from baseline imaging have been developed for extranodal natural killer/T-cell lymphoma (ENKTL) using PET/CT (AUROC 0.88) and for mantle cell lymphoma (MCL) using CT-based radiomic and deep features (AUROC 0.70). For both conditions, external cohort validation is still lacking, which is a critical gap since relapse prediction models need to generalize across different imaging platforms and patient populations to be clinically actionable. These advanced-phase applications represent the frontier of DL in hematology and will require substantially larger datasets and prospective validation before informing treatment decisions.
Validation gaps: The most pervasive limitation across the 65 reviewed manuscripts is the predominance of internal validation. The majority of studies used internal cross-validation strategies, and only a minority were evaluated on external cohorts from separate institutions. For the few studies that did apply external validation, performance consistently dropped compared to internal results, confirming that overfitting to training data characteristics (patient demographics, staining protocols, scanner types, institutional acquisition practices) is a real and recurring problem. Prospective validation cohorts, representing the highest evidence standard, were present in only one of the 65 reviewed manuscripts.
Explainability: DL models are frequently criticized as "black boxes" that produce predictions without interpretable reasoning. The review documents that explainability methods were used in only a subset of studies, and even when applied, they do not necessarily validate biological plausibility. Methods such as Grad-CAM, SmoothGrad, and attention heatmaps highlight which image regions drove a prediction, while UMAP and t-SNE visualize how the model clusters cases in high-dimensional feature space. However, none of these methods provide mechanistic guarantees, and multiple examples from other domains demonstrate that seemingly high-performing models exploit artifactual features rather than biologically meaningful ones. Uncertainty quantification, which would flag low-confidence predictions for human review, is rarely reported despite being arguably necessary for clinical deployment.
Human comparison and clinical integration: Direct comparison of DL model performance against expert clinicians (hematopathologists, radiologists) was performed in only a small number of the reviewed studies. Without this comparison it is difficult to quantify the clinical benefit of AI assistance. Furthermore, even well-performing models face structural barriers to clinical integration: there are no large publicly annotated multi-modal datasets for most hematologic conditions analogous to The Cancer Genome Atlas for solid tumors, critical morphological detail in hematopathology may require magnification levels beyond standard pathology scanners, and standardized reporting guidelines for AI in hematology (analogous to TRIPOD or CONSORT-AI) are still in development.
Future priorities: The authors identify several high-priority directions. Multi-scale image analysis is needed because hematopathologists assess specimens at multiple magnification levels, yet most current DL models process a single resolution. Multi-modal architectures integrating histopathology images with flow cytometry, molecular profiling, cytogenetics, and clinical variables represent the next logical step. DL applications are notably absent from stem cell transplantation workflows and from thrombotic microangiopathies where morphological assessment is critical. Standardization of DL protocol design under frameworks such as SPIRIT-AI, STARD-AI, and CONSORT-AI will be necessary to enable meaningful cross-study comparison and to support regulatory review for clinical deployment.