Examining bone marrow is a fundamental part of diagnosing and monitoring blood cancers and other hematologic diseases. A key procedure is the differential cell count (DCC), in which a trained pathologist looks at a bone marrow aspirate smear under a microscope and manually counts the proportions of different cell types present.
DCCs are clinically critical because disease definitions are based on these proportions. For example, acute myeloid leukemia (AML) is diagnosed when myeloblasts (immature malignant cells) comprise 20% or more of bone marrow cells. Multiple myeloma is defined by elevated percentages of plasma cells. Even a small error in counting can push a patient above or below a diagnostic threshold, directly affecting treatment decisions.
Manual DCCs have significant limitations: they are labor-intensive and time-consuming, require highly trained pathologists, and are subject to inter- and intra-observer variability - different experts looking at the same slide may reach different counts. The number of cells counted is also typically limited to 300-500, introducing statistical sampling error. Automation could eliminate most of these problems.
Unlike standard blood tests, where cells flow through analyzers in dilute, well-separated suspension, bone marrow aspirate smears contain densely packed, touching, and overlapping cells with highly diverse morphologies. Cells from different stages of development look similar, and distinguishing them requires subtle assessment of nuclear shape, chromatin texture, and cytoplasm color.
Traditional automated analyzers were designed for peripheral blood, not bone marrow, and have failed to handle the complexity of bone marrow specimens reliably. Common failures include misidentifying multi-lobed nuclei as multiple separate cells, failing to count nucleated red blood cells, and being confused by fat droplets (lipid) in the marrow.
Deep learning, particularly convolutional neural networks (CNNs), offers a fundamentally different approach. Rather than applying rigid, hand-coded rules about cell appearance, CNNs learn directly from labeled examples, adapting to the natural variability in staining and morphology. However, CNNs require thousands of accurately labeled training examples - a bottleneck the researchers had to solve first.
The team scanned 17 non-neoplastic (normal) bone marrow aspirate smears and 5 disease cases (3 AML, 2 multiple myeloma) at 40x magnification (0.25 micrometers per pixel) using a whole-slide imaging scanner. The resulting digital images are extremely large - essentially digital pathology slides that can be viewed and analyzed on a computer just like a physical microscope slide.
To annotate the training data, the team used a web-based Digital Slide Archive (DSA) platform that allows pathologists to view, annotate, and label cells in high-resolution digital slides. Pathologists placed point annotations at the center of each cell and assigned it to one of 13 cytological classes (including myeloblasts, promyelocytes, myelocytes, lymphocytes, monocytes, plasma cells, erythroid precursors, and others). Students then drew bounding boxes around each pre-identified cell for training the detection algorithm.
This tiered approach was efficient: the expert pathologists did the cognitively demanding work of cell classification using a simple point-click interface, while students performed the more mechanical work of drawing bounding boxes. The result was over 10,000 annotated cells - a scale critical for training high-performance CNNs. All annotations were reviewed by three pathologists in a consensus session to correct any errors.
The system uses two separate, sequential neural networks. The first is a detection network based on Faster R-CNN (Region-Based Convolutional Neural Network) - an architecture that identifies where cells are located in an image and draws bounding boxes around them. Crucially, at this stage all cells are treated identically; the detection network just finds 'objects' without caring what type they are.
The second is a classification network based on VGG16, a deep CNN originally developed for natural image recognition. Once a cell has been detected and a small image patch extracted around it, the classifier determines which of the 12 cytological categories it belongs to. Separating detection from classification allows each network to be independently optimized for its specific task.
A key innovation was test-time augmentation for classification: instead of classifying each detected cell once, the system generates 16 augmented versions of each cell image (different rotations, flips, brightness, and contrast adjustments) and aggregates the predictions across all 16 versions. This ensemble approach proved critical, improving overall classification AUC by an average of 5% - a statistically significant gain.
Cell detection performance, measured by precision-recall area under curve (AUC), was 0.959 with a standard deviation of 0.008 across 6-fold cross-validation on the 17 non-neoplastic cases. This means the system reliably found nearly all bone marrow cells in the slide, with very few missed cells or false detections. The median relative positioning error for detected cells was just 6%, ensuring that the cropped images fed to the classifier were well-centered on the actual cells.
Cell classification performance was even more impressive, reaching a median ROC AUC of 0.982 with standard deviation 0.03 across all cytological classes. Individual class AUCs ranged from 0.960 for monocytes (the most challenging) to 1.00 for basophils (perfectly classified). The most common errors were between closely related cell types: lymphocytes misclassified as unknown (21%), monocytes as unknown (14%), and blasts confused with promyelocytes (10%).
Errors between adjacent cell types in the myeloid maturation series (blast, promyelocyte, myelocyte) are understandable and clinically meaningful - these distinctions require assessing subtle nuclear chromatin patterns that challenge even experienced pathologists. The fact that the model's errors follow biologically sensible patterns suggests it has learned genuine morphological features rather than spurious correlations.
A critical test was applying the system to disease cases - 3 AML samples and 2 multiple myeloma samples - despite the classifier having been trained only on non-neoplastic (normal) cells. This tests whether the system can generalize to abnormal, malignant cells that may look somewhat different from their normal counterparts.
Detection performance remained strong: AUC of 0.970 for AML and 0.979 for multiple myeloma cases. Classification of AML blasts achieved an AUC of 0.893, and when cells classified as 'unknown' were counted as correct (which is actually how pathologists handle ambiguous blasts in practice - labeling them unknown until further workup), accuracy rose to 93.3%.
For multiple myeloma, neoplastic plasma cell classification AUC was 0.970, rising to 0.988 when unknowns were counted as correct. These results are particularly encouraging because the classifiers were never trained on malignant cells - they generalized from normal cell morphology alone. This suggests that adding even modest amounts of annotated disease data would substantially improve performance on malignant cases.
The clinical implications of automating bone marrow DCCs are substantial. Currently, pathologists count 300-500 cells per case by hand, a process that takes 15-30 minutes of skilled labor per slide. An automated system could count thousands of cells in a few minutes, dramatically increasing statistical accuracy - especially near the diagnostic thresholds that define disease (such as the 20% blast cutoff for AML).
The system processes an ROI containing 500 cells in approximately 162 seconds on a GPU-equipped server. This is fast enough for clinical workflow, especially since processing can happen in the background while slides are being prepared or reviewed by a pathologist for other purposes.
Standardization is another key benefit. Currently, DCC results vary between institutions and even between pathologists at the same institution. An automated system that applies the same classification rules consistently would reduce this variability, making comparative studies across hospitals more meaningful and potentially improving diagnostic agreement for borderline cases.
This study represents an important proof-of-concept for machine learning-based bone marrow DCC automation. The researchers successfully built a large annotated training dataset, developed an effective two-stage detection-classification pipeline, and demonstrated compelling accuracy on both normal and disease samples.
Key limitations acknowledged by the authors include the small number of disease cases tested (3 AML, 2 myeloma) and the fact that the images were acquired at 40x rather than the 100x routinely used for clinical diagnosis. The team notes that 40x provides sufficient detail for cell-type identification, but higher magnifications may be needed for detecting finer cytomorphological details critical for some diagnostic distinctions.
Future work will focus on expanding the dataset to include a comprehensive array of bone marrow diseases (not just AML and myeloma), formal clinical validation studies comparing automated DCCs against expert manual counts, and integration into clinical laboratory workflows. If validated at clinical scale, this technology could fundamentally transform how bone marrow diagnoses are made in hematologic oncology.