Acute myeloid leukemia (AML) is a fast-growing blood cancer in which immature blood cells called myeloblasts multiply uncontrollably. Diagnosing AML requires careful examination of bone marrow smears - thin preparations of bone marrow cells spread on glass slides and stained so that individual cells can be seen under a microscope.
The current standard practice relies on experienced hematopathologists (physicians specialized in blood diseases) to manually count and classify cells. This process is time-consuming, and results can vary depending on which expert performs the analysis - a problem known as inter-observer variability. A more consistent, automated approach is urgently needed.
One of the most clinically important genetic changes in AML is a mutation in the NPM1 gene (Nucleophosmin 1). NPM1 mutations occur in roughly one-third of all adult AML cases and in up to 50-60% of cases with normal chromosomes. Identifying whether a patient has an NPM1 mutation is important because it affects risk stratification - the process of classifying patients into low, intermediate, or high risk groups to guide treatment decisions.
Currently, NPM1 mutation status is determined through separate molecular laboratory tests. If this information could be inferred directly from microscope images, it could speed up diagnosis and reduce the need for additional costly tests.
Researchers at the University Hospital Dresden assembled a dataset of 1,251 AML patients drawn from multiple clinical trials and registries, plus 236 healthy bone marrow donors as controls. High-resolution photographs of bone marrow smears were taken using a powerful microscope at 50-fold magnification.
A multi-step deep learning (DL) pipeline was constructed. Deep learning is a type of artificial intelligence that learns patterns from large amounts of labeled data by passing information through many layers of mathematical processing units, called artificial neurons. The first step used a Faster Region-based Convolutional Neural Network (FRCNN) to automatically identify and outline individual cells in each image - a process called cell segmentation.
A human-in-the-loop approach was used during segmentation: hematologists first manually corrected the AI's cell border proposals, and those corrections were fed back to improve the model. After enough training cycles, the AI could segment new images without human intervention. In total, over 94,000 individual cells were manually segmented to train this system.
For the AML versus healthy classification task, an Ensemble Neural Network (ENN) combined the outputs of multiple models trained on different cell features such as blast ratios, presence of Auer rods (abnormal cellular structures), eccentricity, and color. For predicting NPM1 mutation status, a separate model called ResNet50 - pre-trained on a large image database called ImageNet - was fine-tuned using the bone marrow slide images.
The AML detection model achieved an area under the receiver operating characteristic curve (AUROC) of 0.9699. AUROC is a standard measure of diagnostic accuracy, where 1.0 is perfect and 0.5 is no better than chance. A score near 0.97 indicates the model distinguishes AML from healthy samples with very high reliability.
The model correctly identified 87% of AML patients and 89% of healthy controls in the validation set - the withheld portion of data the model had never seen during training. Misclassifications were relatively low: 13% of AML cases were incorrectly labeled as healthy, and 11% of healthy donors were incorrectly labeled as AML.
A key innovation was the multistage approach: combining individual models for each cell feature into an ensemble performed much better than using a single model on whole images alone. Testing a simpler approach of classifying disease from single cells (rather than aggregated smear data) yielded only an AUROC of 0.53 - barely above chance - demonstrating that the context of the full slide matters greatly.
The researchers applied image augmentation techniques - such as mirroring, rotating, and adjusting brightness - to artificially expand their dataset and prevent the model from overfitting (memorizing specific training examples rather than learning general patterns). This helped achieve more robust performance on new data.
Beyond detecting AML, the deep learning model was also tested for its ability to predict NPM1 mutation status using only microscope images - with no molecular testing. The model achieved an accuracy of 86% and an AUROC of 0.92, indicating strong predictive capability for distinguishing NPM1-mutated from NPM1 wild-type AML.
To understand which visual features drove the model's predictions, the researchers used occlusion sensitivity maps. This technique works by systematically blocking different parts of an image from the model and measuring how much its accuracy drops. Areas that, when hidden, greatly reduce accuracy must be important to the prediction - and these are highlighted in a heatmap.
The analysis revealed that the model relied on previously unreported morphological features: in NPM1-mutated cases, it focused on condensed chromatin and perinuclear lightening zones in the myeloblasts - subtle patterns in how the genetic material inside the nucleus is organized. In NPM1 wild-type cases, the model focused on prominent nucleoli (structures within the nucleus where ribosomes are made), which were rarely seen in mutated cases.
These visual differences have a plausible biological explanation: the normal NPM1 protein is anchored in the nucleolus, but the mutated form is abnormally exported to the cytoplasm. This translocation may alter nucleolar appearance in ways detectable by AI even before a pathologist could recognize a pattern.
The model struggled more with NPM1-mutated cases that had a lower variant allele frequency (VAF) - the proportion of tumor DNA that carries the mutation. Cases with lower VAF were more likely to be missed (false negatives), suggesting that when the mutation is present in fewer cells, the morphological signal is weaker and harder for the AI to detect.
Among the incorrectly classified NPM1-mutated cases (false negatives), there was a significantly higher proportion of therapy-associated AML (tAML) - cases that arose as a complication of prior cancer treatment. These cases may have distinct cellular morphology that diverges from typical NPM1-mutated AML, confusing the classifier.
False negatives also had lower white blood cell counts and fewer blasts in peripheral blood, suggesting these cases were at an earlier or less proliferative stage where mutation-associated morphological features were not as pronounced. These findings help researchers understand when the model is more or less reliable.
In a notable quality check, the NPM1 classifier - which was never trained on healthy donor samples - was applied to healthy bone marrow images anyway. It correctly labeled 91% of healthy donors as NPM1 wild type, with only 9% false positives. This confirms that the model is not simply learning broad features of AML but is genuinely picking up mutation-specific cellular characteristics.
This AI system could be deployed in clinical laboratories to rapidly flag suspected AML cases as soon as bone marrow images are uploaded, even before results from flow cytometry, cytogenetics, or molecular genetics are available. This could accelerate diagnosis in time-sensitive cases and allow specialists to prioritize urgent workups.
A key limitation is that representative regions of bone marrow slides still needed to be manually selected by a hematologist before being analyzed by the AI. Future work aims to implement whole-slide imaging with automated region-of-interest selection, removing this manual bottleneck entirely.
There is also a potential age-related bias: healthy bone marrow donors were substantially younger than AML patients in this study. Since bone marrow changes with age - becoming less cellular and more fatty - a classifier trained on young healthy donors may behave differently when compared to older healthy individuals. Further validation with age-matched controls is needed.
The study was conducted at a single center. To be used broadly, such models must be trained and validated on data from multiple hospitals, which differ in staining protocols, microscopes, and image processing methods. The authors call for an open-source bone marrow cell database analogous to The Cancer Genome Atlas to facilitate such multi-center AI development.
This study serves as a proof-of-concept that deep learning can extract genetic information from morphology - demonstrating that the visual appearance of cells under a microscope contains information about the underlying molecular biology, even for mutations not previously thought to have visible morphological correlates.
The discovery of novel morphological features - such as condensed chromatin patterns and perinuclear lightening zones - associated with NPM1 mutations may stimulate new biological research into how this mutation physically alters nuclear architecture. AI-driven discovery of cellular features is an exciting direction for future pathology research.
Looking ahead, the researchers envision extending this approach to other clinically important AML mutations such as FLT3, IDH1/2, and DNMT3A. If the visual imprint of these mutations can also be learned by AI, it may become possible to rapidly screen for multiple genetic alterations simultaneously from a single microscope image.
Ultimately, integrating AI image analysis with other diagnostic modalities - such as flow cytometry, cytogenetics, and clinical data - through ensemble learning could produce a comprehensive automated diagnostic tool. Such a tool could reduce diagnostic delays, standardize reporting, and support pathologists in low-resource settings where expert hematopathologists are scarce.