Researchers have developed ALLocate, a novel, low-cost artificial intelligence-powered microscope plugin designed to automate the detection of acute myeloid leukemia (AML) in bone marrow aspirate smears. The system, which integrates a custom-built, 3D-printed automated scanning platform with advanced deep learning models, aims to address the significant diagnostic gap in resource-limited settings where conventional whole-slide scanners—often costing upwards of $300,000 USD—are unavailable. By automating the identification of diagnostic regions and the quantification of leukemic blasts, the system seeks to augment the capacity of clinical laboratories and improve the speed and standardization of hematologic diagnoses without requiring expensive infrastructure.
The ALLocate system operates by first transforming conventional microscope-captured smears into digital images using a motorized stage controlled by an Arduino-based system. This hardware module, which costs approximately $150 to implement, utilizes stepper motors mounted in a 3D-printed shell to interface with the microscope’s manual control knobs, enabling systematic, unidirectional raster scanning of the slides. Once the images are captured, the AI pipeline employs two specialized deep learning modules: a convolutional neural network (CNN) trained to classify regions of interest as either adequate for diagnosis, clotted, or blood-enriched, and a YOLOv8-based model designed to detect and classify individual nucleated cells. This two-stage process allows the system to focus computational resources on diagnostically relevant regions, mimicking the workflow of a human hematopathologist.
To train and validate the models, the researchers utilized a large-scale dataset encompassing more than 11,000 annotated regions and 130,000 annotated cells. The system’s performance was evaluated at three levels: region-of-interest identification, blast cell detection, and end-to-end slide-level diagnosis. The region classifier demonstrated high accuracy, achieving an area under the receiver operating characteristic curve (AUROC) of 0.992 for identifying adequate diagnostic regions. The cell detection module, optimized using a YOLOv8 architecture, achieved a precision, recall, and mean average precision (mAP50) of 0.84, 0.85, and 0.90, respectively. These performance metrics indicate that the system can reliably distinguish between blast cells and normal cells, a critical requirement for accurate leukemia diagnosis.
The study conducted extensive validation using independent multi-institutional cohorts from the University of California, San Francisco (UCSF), and Memorial Sloan Kettering Cancer Center (MSKCC). In an end-to-end evaluation using 40 digitized slides, the model achieved high accuracy in distinguishing AML from normal cases, reaching a near-perfect AUROC with as few as five analyzed regions per slide. Furthermore, the complete ALLocate pipeline was applied to 165 physical glass slides in a real-world clinical setting. On these slides, the system demonstrated robust performance, correctly classifying 94.4% of AML cases and 100% of normal cases in the internal UCSF cohort, and 94.6% of AML cases and 79.5% of normal cases in the external MSKCC cohort. The overall diagnostic performance across these diverse cohorts underscores the system’s potential as a generalizable tool.
A notable advantage of the ALLocate framework is its computational efficiency, with both the region classification and object detection modules lightweight enough to operate entirely on a standard CPU. This capability allows the system to be deployed in environments where high-end graphical processing hardware is unavailable. In clinical comparisons, the system matched the performance of human pathologists in quantifying blast counts while drastically reducing the time required for evaluation. Whereas manual blast counting by a pathologist in 200 regions can take over four hours, the automated system can perform the same analysis in less than one minute, significantly improving laboratory throughput and potentially serving as a rapid triage tool for prioritizing high-risk cases.
Despite these promising results, the researchers identified several limitations that necessitate further development. The current hardware prototype lacks an automated Z-axis focusing mechanism, leading to potential focal variability across different regions of a smear. Additionally, while the system effectively distinguishes AML from normal marrow, it has not been explicitly validated against other hematologic or metastatic conditions. Furthermore, the researchers observed that immature granulocytes with prominent nucleoli can occasionally be misclassified as leukemic blasts, a challenge that may be mitigated in future iterations by incorporating a more diverse range of reactive marrow samples into the training dataset. The current study also emphasizes that ALLocate is designed as an assistive tool rather than a standalone diagnostic replacement, with outputs requiring expert clinical verification.
Future work will focus on integrating automated autofocus functionality and developing a unified graphical user interface to streamline case management and pathologist review. The research team also highlights the broader need for standardized, publicly available bone marrow aspirate datasets with blast-level bounding box annotations to facilitate benchmarking and future model development. Because the current dataset, while large, lacks subtype-balanced cohorts, continued research is essential to fully characterize performance across the full spectrum of AML variants. Nevertheless, the successful integration of low-cost hardware with efficient deep learning models provides a scalable solution for automated marrow smear analysis.
This study represents a significant step toward bridging the gap between cutting-edge AI innovation and practical deployment in resource-limited clinical settings. By reducing the reliance on prohibitively expensive whole-slide scanning systems, the ALLocate framework offers a viable path for increasing diagnostic accessibility for hematologic malignancies. The system’s strong performance on independent, external datasets validates its reliability across institutional workflows and diverse slide preparations. Ultimately, as an assistive tool, ALLocate holds the potential to augment the capacity of overburdened clinical specialists, enhance the standardization of diagnostic workflows, and facilitate timely leukemia detection in under-resourced medical facilities globally.
The global burden of acute myeloid leukemia highlights the clinical need for automated microscopic screening. Acute myeloid leukemia is an aggressive hematologic malignancy caused by the proliferation of hematopoietic stem cells and accounts for roughly 28 percent of leukemia-related deaths worldwide. Formal diagnosis requires identifying at least 20 percent leukemic blasts in marrow or peripheral blood, demanding manual light-microscopic review of 200 to 500 cells per smear. In resource-constrained regions, specialists are scarce; sub-Saharan African nations have only 0.1 to 1.3 hematopathologists per million people. While earlier low-cost automated microscopes priced between $250 and $500 demonstrated success detecting bloodborne parasites such as malaria, Loa loa, and Trypanosoma brucei, they were not engineered for marrow leukocyte differential analysis. ALLocate addresses this specific diagnostic gap.
The physical scanner integrates accessible electromechanical hardware directly onto existing clinical microscopes without modifying underlying optical pathways. Tested on an Olympus BX40 microscope equipped with an Olympus EP50 camera, the scanning system employs two NEMA 14 stepper motors with a 1.8-degree step angle and 0.4-ampere rated current, driven by A4988 motor drivers powered by a 12-volt, 10-ampere supply. An Arduino Mega 2560 board paired with a RepRap Arduino Mega Pololu Shield executes motor control via an Arduino C++ script. The X-axis motor connects directly to the manual stage translation shaft, while the coaxial Y-axis knob interfaces through a geared coupling in a 3D-printed housing guided by steel rods. This apparatus performs systematic unidirectional raster scanning across smears, costing $150 alone or $1,150 including the microscope.
The software pipeline divides automated assessment into sequential deep learning stages. The initial convolutional neural network categorizes smear tiles into adequate diagnostic areas, clotted regions, or red blood cell regions. Implemented in PyTorch with three convolutional and max pooling layers alongside three fully connected layers, the network was trained with cross-entropy loss and the Adam optimizer over 10 epochs at a learning rate of 0.0005, utilizing oversampling on clotted regions. Beyond high accuracy on adequate regions, it attained receiver operating characteristic areas under the curve of 0.998 for blood regions and 0.997 for clotted areas. For downstream cell analysis, the YOLOv8 detector incorporated an explicit artifact category during training alongside normal and blast labels to prevent false positive identifications, operating at an optimized confidence threshold of 0.25.
To evaluate algorithmic efficacy, researchers contextualized ALLocate against previously reported automated marrow classifiers. Prior studies utilized large datasets but often required complex multi-stage architectures; for example, Matek and colleagues trained a ResNeXt-50 on 171,374 expert-annotated cells across 945 patients, reporting 79 percent precision for blast identification, whereas Tayebi and colleagues achieved 84 percent mAP50 using YOLOv4. ALLocate reached 90 percent mAP50 using a computationally lean single-stage detector. Model development drew upon marrow smears reflecting substantial biological heterogeneity, including acute myeloid leukemia with NPM1 mutations, RUNX1::RUNX1T1 fusions, CBFB::MYH11 fusions, KMT2A rearrangements, TP53 mutations, therapy-related disease, myelodysplasia-related changes, and transformations from myeloproliferative neoplasms. Across five cross-validation folds, blast detection yielded tight ninety-five percent confidence intervals: 0.805 to 0.821 for precision, 0.816 to 0.844 for recall, and 0.856 to 0.870 for mAP50.
Independent cohort evaluations demonstrated robust statistical agreement between automated counts and expert hematopathologist assessments. On an external region-level testing cohort of 186 tiles from Memorial Sloan Kettering Cancer Center, ALLocate achieved an AUROC of 0.9012, closely mirroring the expert pathologist benchmark of 0.9359. Paired statistical testing showed no significant difference between automated counts and manual review in normal regions (mean 0.75 versus 0.60 blasts, P = 0.07) or leukemic regions (mean 10.54 versus 11.13 blasts, P = 0.12). Furthermore, on physical glass slides, estimated blast proportions differed decisively between diseased and benign marrow across both the internal cohort (mean 40.28 percent versus 0.41 percent, P = 4.72 × 10⁻¹⁰) and the external cohort (mean 51.86 percent versus 4.04 percent, P = 1.77 × 10⁻²³).
Subject of Research: Development and validation of a low-cost, AI-driven microscopy framework for automated acute myeloid leukemia screening.
Article Title: An AI-powered self-driving microscope for low-cost acute leukemia detection
Article References: Yan, E. S., Sun, S., Yin, Z., Kamali, A., Van Cleave, J., Isgor, I. S., Fried, B., Colorado-Jimenez, C., Dilip, D., Baik, J., Manzo, A., Paulsen, S., Singi, S., Drapeau, S., Eren, O. C., Chun, J. K., Syed, A., Cardillo, A., Pulitzer, M., … Goldgof, G. M. (2026). An AI-powered self-driving microscope for low-cost acute leukemia detection. Nature Communications, 17(1), Article 10304. https://doi.org/10.1038/s41467-026-77130-x
Image Credits: AI Generated
DOI: 10.1038/s41467-026-77130-x
Keywords: acute myeloid leukemia, automated microscopy, computational pathology, deep learning, bone marrow aspirate, YOLOv8, diagnostic screening, resource-limited settings, clinical decision support, blast quantification, artificial intelligence in medicine, medical hardware development, hematopathology, digital pathology, cost-effective diagnostics
News Source: Nathaniel Bowman. (November 16, 2026). Low-Cost Automated Microscope Plugin Enables Rapid Leukemia Detection. Scienmag.



