Diagnosing acute myeloid leukemia from the bone-marrow smear: a "cell-to-patient" pipeline that sidesteps blast counting (arXiv, 2026)

Posted on 9 June 2026 on arXiv, this preprint proposes a deep-learning pipeline that reads a bone-marrow smear cell by cell, then aggregates those hundreds of observations into a per-patient score to support the diagnosis of acute myeloid leukemia (AML). Trained and validated on 258 patients from six centers — 89 of them held out for external validation — it holds a weighted F1 of 0.87 to 0.91 on three centers never seen during training. The work tackles the right link, the step from cell to patient that most studies ignore, and tests it honestly off its home turf. But the target the model learns is a composite morphological category, explicitly distinct from the leukemic blast, and the preprint reports no patient-level diagnostic sensitivity or specificity, no code, and no comparison to a human cytologist.

The context

Acute myeloid leukemia is a blood cancer in which immature cells of the myeloid lineage, the blasts, flood the bone marrow and crowd out normal blood-cell production. Its diagnosis rests historically on the bone-marrow smear: marrow spread on a slide, stained, then examined under the microscope by a cytologist who identifies, counts and classifies hundreds of cells to estimate, in particular, the percentage of blasts. The reference classifications (WHO, ICC) classically use a threshold around 20% blasts — modulated by genetics — to make the diagnosis.

That manual count is slow, tiring and subject to marked inter-observer variability. Hence a long line of AI work that learns to classify an isolated marrow cell from public datasets. The trouble is that classifying a cell correctly does not tell you how to diagnose a patient: one has to aggregate hundreds of cell-level verdicts into a single decision, and to do so on slides from other hospitals, other stains, other microscopes. It is precisely this link — the cell-to-patient step and its generalization — that this preprint sets out to handle end to end.

The method

The work, authored by Yuqi Ma, Tianyi Wang, Xiaodong Mo, Gen Yang and colleagues, gathers 258 patients from six anonymized centers: a main cohort of 169 patients (centers 1 to 3) for development, and an external validation cohort of 89 patients (centers 4 to 6) never seen during training. Cells are described by a 16-category annotation vocabulary spanning the major lineages (granulocytic, monocytic, erythroid, lymphoid, eosinophilic, and so on).

A central and debatable choice: rather than targeting the leukemic blast in the strict diagnostic sense — hard to annotate reproducibly — the authors define a composite target, the CBLC (Composite Blast-like Cells), a grouping of eight morphological types (N, N1, M, M1, R, R1, J, J1) following a project-wide standard. The model thus learns to spot cells that resemble blasts, not the blast as defined by clinical criteria.

The pipeline has two stages. First a frozen YOLO segmentation module ("You Only Look Once," a family of networks that locate objects in an image in a single pass) detects the cells; its predicted contours are matched to expert-drawn polygons via a contour IoU (Intersection over Union measures the overlap between two shapes, from 0 to 1), and standardized single-cell crops are cut out. Then an EfficientNet-B0 classifier (a compact, frugal convolutional network) assigns a category to each crop. Its training follows a two-stage strategy, GT-to-YOLO then YOLO-to-YOLO: the classifier is first learned on crops from the expert ground truth, then re-adapted on the crops actually produced by YOLO — so as to match inference conditions rather than perfectly outlined cells. Added to this are class-imbalance correction, a "center-border" regularization (penalizing what depends on the cell's edges rather than its center) and morphology-assisted supervision. The cell-level verdicts are finally aggregated into a per-patient CBLC ratio, which serves as support for AML-oriented diagnosis.

The results

On external validation, the pipeline (as an ensemble) reaches a weighted F1 of 0.9076 at center 4, 0.8696 at center 5 and 0.9124 at center 6. F1 is the harmonic mean of precision and recall; "weighted" means each category counts in proportion to its size. Internal validation is described as "stable," with no detailed figure in the abstract. Performance therefore holds at a comparable level across three unseen centers, which is the paper's strong argument.

Clinical translation: what these numbers measure is the quality of cell sorting — the ability to label a slide's cells correctly, center by center. As they stand, they say nothing about final diagnostic performance: the abstract reports no sensitivity, no specificity, no predictive value at the patient level, nor the CBLC-ratio threshold that would tip a patient toward "probable AML." We know how to sort cells; we still do not know how reliably we would diagnose a patient.

What is good

The right problem, tackled end to end. Where most work stops at classifying an isolated cell, this pipeline explicitly handles cell-to-patient aggregation and builds a per-patient score. That is the genuinely useful link, too often skipped, between a lab demonstration and a diagnostic aid.

External validation already at the preprint stage. Holding out 89 patients from three entirely separate centers, and keeping a weighted F1 around 0.87–0.91 there, is more demanding than a simple internal cross-validation. It is exactly the precaution many of the field's publications lack, and it makes the generalization signal credible.

Realistic training against inference drift. The YOLO-to-YOLO strategy, which re-trains the classifier on the imperfect crops actually produced by the detector rather than on perfectly outlined cells, anticipates the degradation seen in real conditions. It is the kind of methodological hygiene that separates an optimistic prototype from a system built to work off the bench.

What is less good

A target that is a proxy, not the diagnosis (failure mode: the misleading task definition). CBLC is a composite "blast-like" category, defined by the project and explicitly separated from the leukemic blast. Yet AML diagnosis rests on a blast percentage and on genetics. Measuring a CBLC ratio is not measuring a diagnosis, and the link between that ratio and the clinical decision — which threshold, for what sensitivity — is not validated here.

Cell-level, not patient-level metrics (failure mode: the misleading metric). The only released figure is a weighted F1 per cell. Such an F1 is dominated by the abundant classes — the normal cells — so it can stay high even while detection of the rare, decisive CBLC is poor. No patient-level sensitivity or specificity, no ROC curve, and above all no comparison to a human cytologist: we therefore cannot tell whether the tool would do better, as well, or worse than an expert.

Uncertain reproducibility and representativeness (failure modes: reproducibility and population bias). No code or weights are released, and the CC BY-NC-ND 4.0 license (non-commercial, no derivatives) effectively bars exploitation and extension. The data come from non-public "anonymized centers," on a small sample (258 patients, 89 external). Nothing is said about the diversity of staining protocols or microscopes — a classic ground for shortcut learning, where the model learns a stain's signature rather than the cell's biology. Finally, the abstract mentions neither funding nor conflicts of interest.

What it changes

For the research community, the message is twofold: the cell-to-patient step deserves to be treated as a problem in its own right, and external validation should be the norm at the preprint stage, not a later add-on. The idea of a robust composite target (CBLC) rather than the hard-to-annotate strict blast is reusable for getting around the rarity and ambiguity of annotations — provided its real diagnostic value is then measured.

For clinicians, nothing changes in the lab today. This is a preprint not yet peer-reviewed, with no patient-level diagnostic metric and no comparison to cytologists. The useful reflex, faced with this kind of tool, is to demand per-patient sensitivity and specificity and the chosen decision threshold, not a cell-level F1 that flatters appearances.

For patients and the public, the lesson is measured: automating the marrow smear is genuinely advancing, but "sorting cells well" is not yet "diagnosing leukemia well." The promise of a fast, consistent digital second reader stays credible; its value will be judged on studies that measure the diagnosis made, not merely the quality of the sort.

Going further

The full preprint is available on arXiv:2606.10735 (DOI 10.48550/arXiv.2606.10735), posted 9 June 2026, under a CC BY-NC-ND 4.0 license, listed under cs.CV and physics.med-ph, with four figures. For the diagnostic framework of AML and the role of the blast percentage, the criteria of the WHO classification of haematopoietic tumours and of the International Consensus Classification (ICC) remain the reference; for AI-based marrow-cell classification, the public bone-marrow cytology datasets (Matek and colleagues) offer a point of comparison.