BMSNet: An AI system for bone marrow smear analysis in leukemia diagnosis

JMIR medical informatics 2020 AI 7 Explanations View Original
Original Paper (PDF)

Unable to display PDF. Download it here or view on PMC.

Plain-English Explanations
Page 1
Bone Marrow Smears in Leukemia Diagnosis

Diagnosing leukemia requires examining bone marrow smears - thin layers of bone marrow cells spread on glass slides and viewed under a microscope. Hematologists count different cell types and identify abnormal cells called blasts, which are immature blood cells that fail to mature normally and accumulate in leukemia.

The proportion of blasts in the bone marrow is a critical diagnostic threshold. By convention, a blast count above 20% confirms acute leukemia. Accurate counting is therefore not just descriptive - it directly determines whether a patient receives a leukemia diagnosis and begins treatment.

Manual bone marrow smear analysis is time-consuming, labor-intensive, and subject to inter-observer variability. Different hematologists may count the same slide differently, and high-quality analysis requires years of specialized training. In many regions, access to experienced hematopathologists is limited.

TL;DR: Manual bone marrow smear analysis for leukemia diagnosis is accurate but slow, subjective, and requires specialized expertise that is not universally available.
Pages 1-2
BMSNet: Deep Learning for Cell Classification

BMSNet (Bone Marrow Smear Network) is a deep learning system designed to automatically classify cells in bone marrow smear images and estimate the blast percentage. It uses YOLO v3 architecture, a state-of-the-art object detection framework originally developed for real-time visual recognition tasks.

The system was trained to recognize 8 categories of bone marrow cells, including blast cells, neutrophil precursors, lymphocytes, monocytes, erythroid precursors, plasma cells, megakaryocytes, and other cells. This multi-class approach captures the full complexity of bone marrow cytology rather than simply detecting blasts in isolation.

A key architectural feature is the use of SE-ResNeXt as the feature extraction backbone. SE-ResNeXt is a deep neural network that learns hierarchical visual features - from simple edges and textures to complex cellular shapes - with channel-wise attention mechanisms that help the model focus on the most discriminative visual elements for each cell type.

TL;DR: BMSNet is a deep learning system using YOLO v3 architecture to automatically classify 8 bone marrow cell types and estimate blast percentage from microscope images.
Pages 2-3
Training Data and Annotation

The system was trained on 17,319 manually annotated cells drawn from 291 high-resolution bone marrow photographs from real patient samples. Each cell was labeled by experienced hematologists, providing the ground truth that the neural network learned to reproduce.

High-resolution imaging is essential for this task because subtle morphological differences between cell types - nuclear shape, chromatin texture, cytoplasm color, and granularity - are the visual cues hematologists use for classification. BMSNet was trained on whole-slide images rather than pre-cropped individual cells, matching how a clinician would view the smear.

The training process used data augmentation - artificially expanding the training set by rotating, flipping, and adjusting brightness of existing images. This technique helps the model generalize to new images that may be stained slightly differently or captured at different magnifications, improving robustness across clinical settings.

TL;DR: BMSNet was trained on 17,319 expert-annotated cells from 291 patient bone marrow photographs, using data augmentation to improve generalization to new samples.
Pages 3-4
Accuracy Versus Expert Hematologists

For the critical clinical task of detecting whether blast cells exceeded the 5% threshold (a common early leukemia indicator), BMSNet achieved an AUC of 0.948 compared to expert hematologists' AUC of 0.929. AUC (area under the receiver operating characteristic curve) measures a classifier's ability to distinguish between positive and negative cases across all thresholds, with 1.0 being perfect.

The fact that BMSNet slightly outperformed senior hematologists on this task is noteworthy. It does not mean the AI replaces clinical expertise - expert judgment incorporates clinical context, patient history, and other factors. Rather, it demonstrates that deep learning has reached a level of visual pattern recognition comparable to human specialists in this narrow task.

In a separate competition cohort, BMSNet's blast percentage estimates showed a mean correlation of 0.960 with flow cytometry measurements - the gold standard quantitative method for blast counting. A correlation of 1.0 would be perfect agreement, making 0.960 an excellent result for a fully automated image analysis tool.

TL;DR: BMSNet achieved AUC 0.948 for blast detection, slightly exceeding expert hematologists, and correlated 0.960 with gold-standard flow cytometry blast counts.
Pages 4-5
How YOLO v3 Detects Cells

YOLO v3 (You Only Look Once) is an object detection framework that processes an entire image in a single pass through the neural network, simultaneously predicting bounding boxes around objects and their class labels. This contrasts with older two-stage detectors that first propose regions and then classify them - YOLO's single-pass approach is dramatically faster.

For bone marrow analysis, speed matters because a single smear slide may contain thousands of cells. YOLO v3 divides each image into a grid and predicts cell locations and types within each grid cell, using multiple detection scales to capture cells of different sizes. The SE-ResNeXt backbone adds squeeze-and-excitation attention blocks that recalibrate feature map weights, helping the network prioritize features most relevant to each cell category.

The combination of YOLO's real-time detection speed and SE-ResNeXt's powerful feature extraction allows BMSNet to analyze a bone marrow smear in seconds - a task that takes a skilled hematologist 30-60 minutes. This efficiency gain could enable rapid preliminary screening before expert review.

TL;DR: YOLO v3's single-pass object detection combined with SE-ResNeXt feature extraction allows BMSNet to analyze thousands of bone marrow cells in seconds rather than hours.
Pages 5-6
Practical Impact on Leukemia Diagnosis

The primary clinical application is screening assistance: BMSNet processes a bone marrow smear and provides a preliminary blast count and cell classification before a hematologist reviews the case. This could reduce the time to diagnosis, particularly in high-volume clinical settings or when expert review is delayed.

A second application is quality control. Comparing BMSNet's cell counts to a hematologist's manual count could flag significant discrepancies for re-examination. Cases where the automated and manual counts diverge substantially may warrant additional testing, potentially catching rare cell types or unusual morphologies that either the AI or the human reviewer might miss alone.

For resource-limited settings without immediate access to specialized hematopathologists, BMSNet could provide a reliable quantitative assessment that enables earlier clinical decisions while awaiting expert consultation. This democratization of specialized diagnostic capacity is a key motivation for developing AI tools in pathology.

TL;DR: BMSNet can provide rapid preliminary blast counts to assist hematologists, with particular value for high-volume labs and settings with limited specialist access.
Page 6
Limitations and Future Directions

BMSNet was developed and validated on data from specific institutions, and performance could vary with different staining protocols, microscope hardware, or patient populations. External validation across diverse clinical sites is a necessary next step before widespread deployment.

The system classifies cells into 8 broad categories, which may not capture all diagnostically relevant subtypes. Future versions could incorporate more granular classification - distinguishing, for example, between myeloblasts and lymphoblasts within the blast category, which has important treatment implications.

Integration with electronic health records and laboratory information systems would be needed for clinical deployment. Regulatory approval processes in different countries also require prospective clinical trials demonstrating improved patient outcomes - a significant but achievable next step given the strong performance already demonstrated.

TL;DR: BMSNet requires external multi-site validation and regulatory review before clinical deployment, with future versions expected to offer more granular cell subtype classification.
Citation: Open Access, 2020. Available at: PMC7177428.