LLD: Large Leukemia Dataset for AI Diagnosis

• LLD is a comprehensive, expert-annotated dataset of microscopy images and clinical metadata for leukemia diagnosis research.
• It integrates diverse imaging modalities, annotations, and clinical data spanning pediatric and adult cohorts across varied specimen types and acquisition protocols.
• LLD supports the development of AI models for cell detection, classification, morphological analysis, and diagnosis prediction using robust evaluation protocols.

The Large Leukemia Dataset (LLD) is a collection of large-scale, expert-annotated microscopy images and metadata designed to support the development of AI methods in leukemia diagnosis. LLD aggregates multimodal imaging data, per-cell and per-patient annotations, and clinical laboratory values across diverse platforms and acquisition protocols. With its breadth—spanning bone marrow aspirate, peripheral blood smears, and both pediatric and adult cohorts—the dataset enables end-to-end studies on leukocyte detection, fine-grained cell classification, morphological attribute analysis, and diagnosis prediction in hematological malignancies.

1. Dataset Scope and Cohort Structure

LLD incorporates data from multiple acquisition sources, reflecting both pediatric and adult populations, as well as different specimen types (bone marrow, peripheral blood film). In its most comprehensive instantiation, as described for bone marrow, the dataset comprises 246 pediatric patients (<18 years) diagnosed with acute lymphoblastic leukemia (ALL), acute myeloid leukemia (AML), or chronic myeloid leukemia (CML). The sample demographics match national incidence (54.2% male, 45.8% female), binned across age intervals [0–3), [4–6), [7–11), [12–15), and [16–18) years. Diagnostic subtypes encompass ten ALL subgroups, nine AML FAB subtypes, and CML phase stratifications. Peripheral blood–derived LLD versions encompass both healthy (WBC subtypes) and malignant (ALL, AML, APML, CLL, CML) cells, with single-cell cutouts, morphological attribute captions, and synthetic domain-shift subsets (Höfener et al., 19 Sep 2025, Rehman et al., 3 Apr 2025, Logtestijn et al., 7 Jan 2026).

2. Imaging, Annotation, and Taxonomies

Imaging protocols in LLD capture diagnostic variability via multi-instrument, multi-magnification acquisition. Bone marrow smears are digitized at 40× magnification (0.11 × 0.11 μm/px) with Pappenheim staining, while peripheral blood samples are prepared at 100×, 40×, and 10×, utilizing both high-cost (Olympus CX23) and low-cost (XSZ-107BN) microscopes, spanning mobile and high-resolution digital camera sensors (Rehman et al., 3 Apr 2025, Rehman et al., 2024).

Annotation is cell-centric and hierarchical. Key statistics from the pediatric bone marrow variant (Höfener et al., 19 Sep 2025):

• 45,176 leukocyte bounding-box annotations across 426 ROIs (regions of interest)
• 28,830 high-quality class labels (33-class taxonomy) over 232 ROIs
• Cell types encompass lineage-specific precursors, blasts, mature granulocytes, erythroid stages, megakaryocytes, and rarities (e.g., Pseudo Gaucher cell, neutrophil extracellular trap)
• Peripheral blood versions catalog 14–33 cell types, with each cell annotated with 6–18 morphological attributes: nuclear chromatin, nuclear shape, nucleoli, cytoplasmic basophilia and vacuolation, cell size, and granularity (Rehman et al., 3 Apr 2025, Rehman et al., 2024, Logtestijn et al., 7 Jan 2026)

Consensus web-based workflows were used for multi-observer labeling, with 87.8% of cells reaching consensus after two annotations. 3.3% required further review. Median per-cell annotation time ranged from 1.2 seconds (validation) to 3.0 seconds (classification view) (Höfener et al., 19 Sep 2025). Expert audits and automated QA pipelines were used for attribute caption quality in vision-language variants (Logtestijn et al., 7 Jan 2026).

3. Metadata, Organization, and Access

LLD standardizes imaging data as DICOM whole-slide images (WSI) with structured reporting records for ROIs and cell-level annotations (bounding boxes, class labels). Extracted single-cell crops are provided in DICOM or standard formats (JPEG/PNG). Patient-linked tabular metadata files (CSV) provide demographics, diagnosis, 18 LOINC-coded laboratory values, and manual differential cell counts (DCC). Data are organized in a directory hierarchy by patient ID, encompassing slide, ROI, cell, and annotation folders, plus a global metadata directory (Höfener et al., 19 Sep 2025).

A representative structure:

/LLD/ /patient_<ID>/ /slide/ /rois/ /cells/ /annotations/ /metadata/

LLD is distributed via respected repositories such as Zenodo and the National Cancer Institute Imaging Data Commons (DOI: 10.5281/zenodo.15490664), licensed under Creative Commons Attribution 4.0 International (CC BY 4.0). No registration is required (Höfener et al., 19 Sep 2025).

4. Data Partitioning and Domain Diversity

Patient-level splits use stratified strategies to preserve diagnosis subtype and cell-class distributions. Bone marrow LLD is partitioned into a 60%/20%/20% split (~148/49/49 patients) for training/validation/test. The distribution ensures proportional representation of abundant and rare cell classes (e.g., Lymphocytic Blast at ~22.6%, Pseudo Gaucher cell at ~0.3%) (Höfener et al., 19 Sep 2025).

Peripheral blood and domain-adaptation studies further stratify data by microscope/camera/magnification "domain," enabling domain-shift analysis. Domain adaptation methods are evaluated by cross-domain testing (e.g., training on high-cost/100×, testing on low-cost/100×), with substantial drops in AP when models are transferred between hardware domains (Rehman et al., 2024, Rehman et al., 3 Apr 2025).

5. Baseline Methods and Evaluation Protocols

LLD is designed to support end-to-end pipelines:

• Cell detection: Trained one-stage detectors (e.g., CenterNet, YOLOv5, FCOS, DINO) and two-stage networks (Faster R-CNN) on ROIs. On bone marrow smears, Faster R-CNN achieved Precision=0.967, Recall=0.945, F1=0.956, AP=0.958 at IoU>0.5 (Höfener et al., 19 Sep 2025). YOLOv5x reached mAP@50=44.2 on peripheral blood domains (Rehman et al., 3 Apr 2025).
• Cell classification: ResNet-50 with dropout and normalized resolution metadata, softmax over 33 classes, macro AUROC=0.981, median F1=0.615, mean F1=0.577 (Höfener et al., 19 Sep 2025).
• Morphological attribute prediction: Multi-task models such as AttriDet (YOLOv5 backbone plus attribute head) use asymmetric multi-label binary cross-entropy. Baseline F1 for key attributes (chromatin, shape): NC=73.9, NS=95.9, N=54.3, C=89.7, CB=83.6, CV=29.1 (Rehman et al., 3 Apr 2025).
• Diagnosis prediction: Gradient boosting (HistGradientBoostingClassifier) on predicted DCC features achieves mean F1=0.90 (ALL, AML, CML) on held-out patients (Höfener et al., 19 Sep 2025).

Key metrics:

• Average Precision: $\mathrm{AP}=\sum_{n}(R_{n}-R_{n-1})P_{n}$, with $P_{n}$ and $R_{n}$ denoting precision and recall at the n-th threshold.
• F1-score: $$F_{1}=2\frac{\mathrm{Precision}\times\mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}}$$
• AUC, Top-N accuracy, and attribute-wise F1 reported for classifier and VLM models.

6. Domain Adaptation and Sparse Annotation Strategies

LLD is structured for evaluation under realistic clinical constraints including annotation sparsity and domain variability. Sparse annotation protocols (Sparse-LeukemiaAttri subset) annotate only 1/5 of each field of view; learning on full-field images is regularized via a composite loss comprising labeled-region, pseudo-label, and similarity-guided triplet components (Rehman et al., 3 Apr 2025).

Unsupervised domain adaptation (UDA) methods (e.g., DACA, ConfMix) are benchmarked, with mAP@50-95 decreasing substantially under domain shift (e.g., YOLOv5: 44.2 to 25.5), highlighting the relevance for real-world deployment (Rehman et al., 2024, Rehman et al., 3 Apr 2025).

7. Specialized LLD Derivatives and Extensions

Specific LLD derivatives target explainable AI and vision-language benchmarking:

• Morphological captioning: HemBLIP constructed a subset (n=14,659 images: 7,037 healthy, 7,622 leukemic) with 18-attribute expert-annotated captions, supporting vision-language modeling and interpretable cell description (Logtestijn et al., 7 Jan 2026).
• C-NMC_2019: A single-cell LLD variant (n=10,661 images; 7,272 ALL, 3,389 normal B-precursors) is used for binary ALL/healthy classification, with CNN accuracy up to 94.3% (Papaioannou et al., 2024).

Peripheral blood datasets (Rehman et al., 2024, Rehman et al., 3 Apr 2025) provide >29,000 images across 12 acquisition domains, 10,300+ cell annotations with 7 morphological attributes at 100×, transferred by homography to 40×/10× levels.

---

LLD’s comprehensive coverage of multi-class hematological pathology, its modular metadata, and domain diversity enable broad research into automated diagnosis, morphological analysis, captioning, and robust deployment under resource and annotation constraints. All reported data statistics, taxonomy definitions, annotation and evaluation workflows, and licensing/access details are provided verbatim from the cited sources (Höfener et al., 19 Sep 2025, Rehman et al., 3 Apr 2025, Papaioannou et al., 2024, Logtestijn et al., 7 Jan 2026, Rehman et al., 2024).