Will AI Replace Pathologists? The Truth from Leading Hematology Labs in 2025

Introduction

Artificial intelligence in pathology is advancing at a remarkable pace, raising the question: Will AI replace pathologists in the coming years? DeepFlow, the world’s first flow cytometry AI cloud diagnosis system, demonstrates an accuracy rate of 95% when diagnosing acute leukemia, working approximately 100 times faster than human doctors. Despite these impressive capabilities, the integration of AI into hematopathology presents a more nuanced reality than complete replacement.

AI has the power to change medicine, especially in fields that deal with images, like pathology. AI has done a great job of looking at solid tumours where patterns are easier to find, but it hasn’t been used much in hematopathology yet. This difference shows how complicated and rare hematological neoplasms are. Nevertheless, experts widely agree that AI will be routinely and impactfully used within anatomic pathology laboratories and pathologists’ clinical workflows by 2030.

The future is one of collaboration rather than replacement. AI-driven algorithms are revolutionizing diagnostics through the automation of tasks like blood smear analysis, cell classification, flow cytometry, and early disease detection. Furthermore, artificial intelligence may be integrated into the morphology, immunophenotype, cytogenetics, and molecular biology workflow throughout sample processing, disease diagnosis, prognosis prediction, and treatment monitoring. These changes point to a complementary relationship between AI and pathologists, where technology enhances human capabilities rather than replacing them.

This article examines the current state of AI in hematopathology, explores leading tools and technologies, analyzes persistent challenges, and presents insights from leading hematology labs about the actual future relationship between AI and pathologists.

How AI is currently used in hematology labs

Artificial intelligence has rapidly transitioned from a theoretical concept to a practical application in real-world blood labs worldwide. Modern hematopathology now utilizes various AI applications to enhance diagnostic accuracy and efficiency, addressing the time-consuming and subjective nature of traditional manual analysis.

Cell classification and counting

AI-powered algorithms now excel at automated cell detection and classification in peripheral blood and bone marrow samples. The YOLO (You Only Look Once) model, which was made to find targets quickly, correctly finds and classifies single cells in bone marrow smears. The YOLOX-s model is one of the newest developments. When used with the MLFL-Net architecture, it gets very high accuracy rates of 89.53% for general classification and 92.5% for predicting acute leukemia.

In 2023, YOLOv7 became a game-changing technology for counting blood cells. It processed images faster than other models for recognition, such as R-CNN or SSD. This progress has greatly improved lab tests by making it possible to detect objects in real time with high accuracy. Also, modern convolutional neural networks (CNNs) are very good at classifying blood cells, with test accuracy reaching 99.12% in recent studies.

Leukemia detection and subtype identification

AI systems have proven remarkably effective at identifying leukemia subtypes. Current research favours either highly accurate single-algorithm systems or hybrid approaches that combine multiple algorithms for early leukemia detection. Several AI models, including artificial neural networks, feed-forward neural networks, and combinations like AlexNet+SVM and ResNet-18+SVM, can achieve 100% accuracy in leukemia diagnosis under controlled conditions.

For chronic myeloid leukemia (CML), the hybrid convolutional neural network method with interactive self-learning school algorithm (HCNN-IAS) achieved perfect 100% accuracy and sensitivity in diagnosing and classifying the disease using blood smear images. Moreover, generative adversarial networks have demonstrated 99.8% accuracy and 98.5% sensitivity in classifying blood smear images as leukemic.

Beyond mere classification, AI systems now reveal morphological features that predict the mutation status of genes in leukemia, including the NPM1 mutation common in acute myeloid leukemia (AML). Digital image-based leukemia diagnosis offers a more straightforward and faster approach than manual examination, eliminating human bias and error while requiring minimal clinical expertise.

Bone marrow smear analysis

Bone marrow cytology remains essential for hematological diagnosis, but traditional manual analysis is labour-intensive and highly subjective. AI-based systems now rapidly and automatically detect suitable regions for cytology and subsequently identify and classify bone marrow cells in each area. This collective cytomorphological information creates a “Histogram of Cell Types” (HCT) that serves as a cytological patient fingerprint.

In a recent validation study, an AI-based algorithm trained on 542 slides containing 597,222 annotated cells achieved an accuracy of 0.94 for the testing dataset. Subsequently, when validated against a multinational real-world dataset comprising 200,639 cells, the AI model maintained an accuracy of 0.881 in classifying individual cells with high precision for blasts (0.927), bands and polymorphonuclear neutrophils (0.955), and plasma cells (0.930).

Flow cytometry automation

Flow cytometry, which is necessary for diagnosing blood diseases, has always needed manual interpretation, which is both time-consuming and error-prone. New AI methods have made this process easier by using advanced visualization and automated gated analysis methods.

DeepFlow, a clinically generalized flow cytometry panel based on multidimensional density-phenotype coupling algorithms, has reduced analysis time to less than 5 minutes per case. In acute leukemia diagnosis, this technology demonstrates 94.6% sensitivity in detecting AML patients and 98.2% sensitivity for B-lymphoblastic leukemia (B-ALL) patients.

Another breakthrough system, the Attune CytPix with automated image analysis, can accomplish in 10 minutes what previously took expert biologists 3 years of full-time study. This technology has completely changed how flow cytometry is done by giving analysts visual proof of cell identity. This breaks the “faith-based flow cytometry” model, in which analysts had to trust that dots on plots represented specific cell types.

Adding AI to hematology labs demonstrates that people are improving their skills rather than being replaced. These technologies enable pathologists to focus on the more challenging aspects of analysis, while AI handles the more routine, time-consuming tasks.

AI in lymph node and bone marrow pathology

Pathological examination of lymph nodes and bone marrow remains a cornerstone of hematological disease diagnosis. In recent years, however, artificial intelligence applications have made substantial inroads into these traditionally human-dominated domains.

Lymph node analysis for B-cell neoplasms

Lymphoid neoplasms, which are tumours that come from immature and mature B lymphocytes, T lymphocytes, and natural killer cells, are very complicated. This makes them an excellent place for AI to help. Deep learning models have shown great promise in lymphoma pathology, with many studies showing that they can make diagnoses more accurate and make it easier to find B-cell neoplasms.

AI algorithms analyze various types of data, including histopathological images, immunohistochemical stains, molecular profiles, and clinical variables, to assist doctors in determining the type of lymphoma a patient has and the most effective treatment. A notable achievement in this area is the application of artificial neural networks to classify mature B-cell neoplasms using gene expression data. In one study, the neural network predicted several lymphoma subtypes with an overall accuracy of 79%, achieving 85% accuracy for follicular lymphoma, 88% for mantle cell lymphoma, 79% for diffuse large B-cell lymphoma, and 80% for Burkitt lymphoma.

The combination of AI and digital pathology platforms has improved diagnostic capabilities through whole-slide imaging (WSI), which allows for remote access, telepathology consultations, and automated image analysis. So, AI algorithms can find quantitative features in WSIs that humans might miss, like cell shape, spatial distribution, and immunohistochemical staining patterns.

Bone marrow evaluation for AML and MDS

Bone marrow evaluation is essential for diagnosing acute myeloid leukemia (AML) and myelodysplastic syndromes (MDS). Currently, MDS diagnosis requires multi-modal procedures including bone marrow cytology, molecular genetics, and cytogenetics—a process both time-consuming and subjective.

In addressing these limitations, researchers have developed several promising AI applications:

Kazemi et al. created a computational framework based on Support Vector Machines (SVM) for classifying AML and its subtypes, achieving 95% sensitivity, 98% specificity, and 96% accuracy in analyzing blood microscopic images.
Lee and colleagues presented a convolutional neural network model that automatically detects dysplasia from bone marrow aspirate images with an excellent area under the curve (AUC) ranging from 0.945 to 0.996.
Rudin et al. demonstrated an AI system for detecting lymphoma infiltration in bone marrow biopsies that achieved 92% sensitivity and 98% specificity, significantly reducing manual examination time.

For MDS detection specifically, researchers have developed deep learning models that analyze bone marrow aspirate smear images with approximately 70% accuracy in internal validation. Another AI pipeline standardized bone marrow smear analysis to address the inherent subjectivity among experts, which can lead to variability in diagnoses.

In essence, machine learning algorithms can predict genetic and cytogenetic aberrations from bone marrow histopathology with impressive accuracy—highest for TET2 mutations (AUC = 0.94), spliceosome mutations (0.89), and chromosome 7 monosomy (0.89). The mutation prediction probability correlates with variant allele frequency, confirming the identification of mutation-specific features.

Challenges in morphological interpretation

Even with these improvements, there are still some big problems that need to be solved before AI can be used in lymph node and bone marrow pathology. First, there are significant problems with the availability and quality of training data. AI algorithms require large, diverse datasets of annotated histopathological images, which consume a considerable amount of resources. To ensure models perform well in clinical settings, it is crucial to verify that the dataset is representative across diverse populations, institutions, and staining protocols

The second significant issue is that AI algorithms are complex to comprehend and explain. Deep learning models can be very complicated and hard to understand, which makes it hard for pathologists to figure out why AI made specific predictions. To build trust and make it easier to use in clinical workflows, it’s important to create models that give clear outputs.

Thirdly, artifacts in histological samples present a fundamental obstacle. Identifying artifacts, particularly in complex specimens like decalcified bone marrow trephines, remains challenging. Issues such as creases, folds, crushing, variations in section thickness, detachment from bony trabeculae, air bubbles, and immunohistochemistry defects can all compromise AI performance.

To date, the primary focus of AI in lymphoma research has been on B-cell neoplasms, particularly follicular lymphoma and diffuse large B-cell lymphoma, which are most frequently encountered in clinical settings. Nevertheless, bone marrow pathology remains an intriguing yet largely untapped area for AI exploration—one that holds promise but requires continued refinement before pathologists can confidently integrate these tools into daily practice.

Tools and models leading the change

Machine learning models form the foundation of AI’s progress in hematopathology, with several key architectures leading this transformation. These tools vary in their approach to data analysis, with each offering distinct advantages for specific diagnostic challenges.

Convolutional Neural Networks (CNNs)

CNNs represent the cornerstone of modern image analysis in hematopathology. Inspired by the visual cortex, these networks process images through layers that identify increasingly complex features—from simple edges to complete cellular structures. In head and neck cancer tissue classification, EfficientNet-B0 models have achieved impressive accuracy rates of 89.9% on previously unseen data.

The architecture of CNNs makes them particularly well-suited for histopathological image analysis. They can identify subtle morphological patterns in tissue through a series of convolutional and pooling layers, followed by fully connected layers. Initially trained on ImageNet (a database of over one million non-medical images), many CNNs in hematopathology employ transfer learning to adapt these pre-trained networks for specialized medical applications.

For classifying blood and bone marrow cells, CNN architectures like GoogLeNet, ResNet-50, and AlexNet have demonstrated exceptional utility. These networks have proven valuable in differentiating tumour growth patterns with high specificity, addressing the histomorphologic heterogeneity that challenges visual microscopic quantification.

Support Vector Machines (SVMs)

Although CNNs dominate modern AI research, SVMs often outperform them in specific hematopathological applications. One study demonstrated that SVM models exceeded CNN-based solutions for blood cell classification by approximately 10%, achieving overall classification rates of about 99%. SVMs with polynomial kernels excel at identifying interpretable features such as cell corners, proportional area, significant axis length, and surrounding relevant area.

For RBC abnormality classification, research suggests SVMs outperform deep learning classifiers because they function effectively with both small and large datasets, whereas deep learning typically requires extensive training data. This versatility makes SVMs particularly valuable in rare hematological conditions where limited samples are available.

The integration of multiple techniques often yields superior results. For instance, a modified U-Net architecture combined with radial basis function-support vector machine (RBF-SVM) demonstrated exceptional accuracy, detecting WBC nuclei with a dice similarity coefficient of 0.972 and recognizing white blood cells with accuracy rates of 99.45%, 98.62%, and 98.81% across three different datasets.

DeepFlow and other commercial tools

Commercial applications now bring these technologies into clinical practice. In 2023, the FDA granted de novo clearance for Scopio Labs’ AI system designed to diagnose blood disorders and cancer. This platform uses high-resolution microscopic imaging to analyze bone marrow biopsies, allowing hemopathologists to evaluate bone marrow smears remotely while increasing diagnostic precision.

DeepFlow, a clinically validated flow cytometry platform using multidimensional density-phenotype coupling algorithms, has demonstrated remarkable diagnostic capabilities. For acute leukemia diagnosis, DeepFlow achieves 94.6% sensitivity in detecting AML patients and 98.2% sensitivity for B-lymphoblastic leukemia patients. Other commercial platforms, such as Attune CytPix with automated image analysis, have dramatically reduced analysis time—accomplishing in minutes what previously required years of expert evaluation.

Performance metrics from real-world studies

The real-world performance of these tools varies across applications. In WBC classification, CNN-SVM hybrid systems achieve up to 99.42% accuracy with 100% sensitivity and specificity on leukemia datasets. For comparing three types of neural networks (MLP, SVM, HRCNN) in classifying white blood cells, studies show MLPs reaching a 99.11% overall correct classification rate, outperforming both SVMs (97.55%) and HRCNNs (88.89%).

When analyzing processing speed, modern AI systems dramatically outpace manual analysis. While traditional methods required 16 minutes for differential counting of 100 white cells, current systems can classify a cell image in less than one second. This extraordinary improvement in processing efficiency, coupled with high accuracy rates, demonstrates why these tools continue to transform hematopathology workflows across clinical settings.

What AI still struggles with in hematopathology

Despite impressive advancements, several critical challenges limit AI’s full potential in hematopathology. These obstacles explain why complete replacement of pathologists remains unlikely in the foreseeable future.

Domain shift and generalizability

Domain shift represents one of the most formidable barriers to AI implementation in clinical settings. Machine learning models perform optimally only when training and test data arise from identical distributions—a condition rarely met in real-world practice. These shifts cause degrading predictive performance over time and produce unreliable performance estimates that can deviate substantially from actual results.

In hematopathology specifically, centre-specific conditions heavily influence pathology data, creating obstacles for widespread adoption. Current pathology domain adaptation methods focus primarily on image patches rather than whole slide images (WSIs), thus failing to capture global features required in typical clinical scenarios. Recent solutions addressing slide-level domain shift have achieved 4.1% AUROC improvement in breast cancer grading cohorts and 3.9% C-index gains in survival prediction.

Artifact detection in histological samples

Artifacts in histological samples present another substantial challenge. Multiple artifacts exist in virtually all digitized slides, related to tissue processing, sectioning, staining, and digitization itself. These imperfections can lead to critical misclassifications, with AI algorithms often failing silently as their architecture cannot typically classify images as “unknown” or “uncertain”.

Common artifacts challenging AI systems include:

Creases, folds, and crushing artifacts
Variations in section thickness
Detachment from bony trabeculae
Air bubbles and immunohistochemistry defects

The presence of histological artifacts can severely hamper the performance of computational pathology systems during automated diagnosis. Unlike human pathologists who intuitively ignore these areas, AI requires sophisticated artifact detection pipelines to exclude irrelevant regions.

Lack of large annotated datasets

The scarcity of well-annotated, publicly available datasets creates a significant bottleneck for developing clinical AI implementations. Labelled pathology datasets are expensive to produce, requiring expert pathologists for precise region-based annotations. This challenge becomes even more pronounced with rare disease types, which account for only a small fraction of available data, causing models to overfit common diseases while performing poorly on underrepresented conditions.

Interpretability of AI decisions

The “black box” nature of many AI algorithms raises concerns about transparency and interpretability, hindering clinician trust and impeding full integration into clinical workflows. Understanding how AI systems arrive at specific conclusions remains challenging—a critical limitation in high-stakes medical settings.

An ML model may base predictions on spurious correlations rather than causal relationships. These correlations may derive from confounding variables like surgical skin markings or hidden variables such as patient age, preparation date, or scanner type. Currently, best-performing AI approaches typically rely on statistical correlations and cannot build causal models to support human understanding.

These challenges collectively underscore why AI currently augments rather than replaces pathologists in hematology labs.

Will AI replace pathologists? What labs are saying

Leading hematology laboratories across the globe are actively implementing AI systems, yet their experiences reveal a more nuanced future than the simple replacement of human experts. Instead, a symbiotic relationship is emerging between artificial intelligence and pathologists.

Insights from leading hematology labs

According to recent consensus studies, experts overwhelmingly agree that AI will be routinely and impactfully used within anatomic pathology laboratory workflows by 2030. Pathologists generally maintain positive attitudes toward AI integration, yet many report a lack of knowledge and skills regarding practical implementation. In response, medical institutions are beginning to incorporate AI education into pathology training programs to prepare specialists for this technological shift.

The growing strain on pathologists worldwide further motivates AI adoption. With demand outstripping supply in many countries, pathology departments face increasing pressure to maintain quality while handling greater workloads. Indeed, laboratory leaders view AI as a potential solution to workforce shortages rather than a replacement threat.

Tasks AI can fully automate

By 2030, several specific pathology tasks are considered highly likely to be fully delegated to AI systems, including:

Verification of positive and negative controls for immunohistochemistry
Prioritization and triage of cases to appropriate pathologists
Slide quality control (detecting tissue folds, tears, stain quality issues)
Screening for microorganisms such as AFB and H. pylori
Screening lymph nodes for metastases
Cervical cytology screening

These automated processes already show promise in pre-screening cervical cytology slides and blood smears, effectively reducing workload for both pathologists and laboratory scientists.

Tasks that still require human expertise

Notwithstanding these advancements, complex diagnostic tasks requiring nuanced interpretation and contextual analysis remain firmly within human expertise. Furthermore, AI systems cannot currently know when they are wrong, necessitating continued human oversight for the foreseeable future. Essentially, “there is no new paradigm where you can implement AI and suddenly the AI takes responsibility for all the decisions it’s making”.

Hybrid models: AI + pathologist collaboration

Rather than replacement, the future points toward augmentation. PathChat, a generalist AI assistant, exemplifies this approach by enabling pathologists to have “conversations” about uploaded images, streamlining workflows by highlighting likely positive cases, ordering additional tests, and drafting reports. Similarly, NaviPath demonstrates how AI can augment pathologists’ capabilities—this human-AI collaborative navigation system helps pathologists identify twice as many relevant patterns in tumour images compared to manual navigation.

Therefore, the consensus among hematology labs suggests AI will enhance rather than replace pathologists, with responsibilities shifting toward technology evaluation, implementation expertise, and more complex interpretative work.

The future of artificial intelligence in hematopathology

The integration of AI into hematopathology is positioned to evolve beyond standalone applications toward sophisticated systems that will reshape diagnostic and therapeutic approaches in the coming years.

Multimodal integration: clinical + molecular + image data

Future AI systems will likely excel through multi-modal integration, combining histopathological images with genomic profiles and electronic health records. Emerging frameworks like graph neural networks permit explicit learning from non-Euclidean relationships, enabling advanced analysis of complex hematopathological data. As part of this shift, researchers envision sophisticated AI assistants that can interpret morphological features of large cohorts while simultaneously processing patient histories.

Predictive modelling for treatment response

AI tools for treatment response prediction represent another frontier. The LORIS system, which integrates tumour mutational burden with five clinical features (patient age, cancer type, treatment history, blood albumin, and inflammation markers), demonstrates superior accuracy in predicting immunotherapy response compared to conventional methods. In multiple myeloma, predictive models have achieved AUC values ranging from 0.78 to 0.90 for treatment response assessment.

Personalized diagnostics and risk stratification

Random Survival Forest algorithms have shown impressive capabilities in risk stratification, achieving c-index values of 0.775 for overall survival prediction in AML patients. Meanwhile, graph neural networks effectively stratify patients’ disease into high and low-risk categories with a hazard ratio of 3.0. These approaches enable treatment personalization based on individual patient characteristics.

Ethical and regulatory considerations

Nonetheless, data privacy, algorithmic transparency, and potential bias remain paramount concerns. Regulatory frameworks continue to evolve, with the FDA developing approaches that balance pre-market assessment for significant changes with streamlined processes for minor algorithm modifications. Besides, explainable AI models addressing the “black box” issue will be crucial for building clinician trust.

Conclusion

Artificial intelligence undoubtedly represents a transformative force in hematopathology, yet the narrative has shifted from replacement to augmentation. Current evidence demonstrates AI’s remarkable capabilities across multiple domains—from achieving 95% accuracy in acute leukemia diagnosis to processing samples 100 times faster than human pathologists. Nevertheless, persistent challenges such as domain shift, artifact detection, and interpretability limitations underscore why complete automation remains elusive.

Leading hematology laboratories worldwide have embraced a collaborative model where AI handles repetitive, time-consuming tasks while pathologists focus on complex interpretive work. This symbiosis allows for enhanced diagnostic accuracy, improved workflow efficiency, and better patient outcomes. Tasks like cell counting, flow cytometry analysis, and preliminary screening can now be delegated to AI systems, freeing pathologists to apply their expertise where human judgment remains essential.

The future certainly points toward increasingly sophisticated AI applications. Multi-modal integration will combine histopathological images with genomic profiles and clinical data, while predictive modelling will enhance treatment response assessment. Random Survival Forest algorithms already achieve impressive c-index values of 0.775 for overall survival prediction in AML patients, demonstrating AI’s potential for personalized medicine.

Though AI continues to advance at an extraordinary pace, the question “Will AI replace pathologists?” misses the point. Instead, we should ask how this technology can best complement human expertise. Pathologists who embrace AI as a powerful tool rather than viewing it as a threat will thrive in this evolving landscape. Their role will undoubtedly grow, shifting toward technology evaluation, implementation expertise, and complex case interpretation.

So, instead of being afraid of becoming obsolete, pathologists should get ready for a future where their jobs change and grow. Because of this, schools need to change their curricula to include AI literacy so that new pathologists are ready to deal with this technological change. AI is still a tool made by people to meet their needs. It is powerful and promising, but it still needs the unique context, judgment, and ethical oversight that only pathologists can provide.

Key Takeaways

Leading hematology labs reveal that AI will augment rather than replace pathologists, creating a collaborative future that enhances diagnostic capabilities while preserving essential human expertise.

AI excels at automating routine tasks like cell counting and flow cytometry, achieving 95% accuracy in acute leukemia diagnosis while processing samples 100 times faster than humans.
Complex diagnostic interpretation, contextual analysis, and ethical oversight remain firmly within human expertise, as AI systems cannot currently know when they’re wrong.
Hybrid AI-pathologist models are emerging as the preferred approach, with AI handling repetitive screening while pathologists focus on nuanced interpretation and complex cases.
Major challenges, including domain shift, artifact detection, and lack of interpretability, prevent full automation, making human oversight essential for reliable clinical implementation.
The future points toward multi-modal AI systems integrating clinical, molecular, and imaging data for personalized diagnostics and treatment response prediction by 2030.

Rather than facing replacement, pathologists should embrace AI as a powerful diagnostic tool that will transform their role toward technology evaluation, implementation expertise, and high-level interpretative work requiring human judgment and contextual understanding.

Frequently Asked Questions:

FAQs

Q1. Will AI completely replace pathologists shortly? No, AI is not expected to replace pathologists completely. Instead, it will augment their capabilities, handling routine tasks while pathologists focus on complex interpretations and oversight—the future points towards a collaborative model between AI and human expertise.

Q2. What tasks can AI currently perform in hematopathology? AI excels at tasks like cell classification and counting, leukemia detection, bone marrow smear analysis, and flow cytometry automation. It can process samples much faster than humans and achieve high accuracy rates in certain diagnostic areas.

Q3. What are the main challenges AI faces in hematopathology? Key challenges include domain shift (difficulty in generalizing across different clinical settings), artifact detection in histological samples, lack of large annotated datasets, and the interpretability of AI decisions. These limitations necessitate continued human oversight.

Q4. How are leading hematology labs integrating AI into their workflows? Leading labs are adopting hybrid models where AI handles repetitive screening tasks while pathologists focus on complex cases and interpretations. This approach aims to improve efficiency and accuracy while maintaining essential human expertise.

Q5. What does the future hold for AI in hematopathology? The future of AI in hematopathology is likely to involve multi-modal integration of clinical, molecular, and imaging data. We can expect more sophisticated predictive modelling for treatment responses and personalized diagnostics. However, ethical considerations and regulatory frameworks will continue to evolve alongside these technological advancements.