WK 1 — Introduction to Computational Hematology

Section 1

What Is Computational Hematology?

Formal Definition

Computational hematology is the application of computational methods — encompassing machine learning, deep learning, statistical modelling, natural language processing, and data engineering — to the acquisition, representation, analysis, and clinical interpretation of haematological data across morphological, cytometric, genomic, and clinical domains.

The field occupies a formally bounded intersection of four parent disciplines. Its outputs are decision-support artefacts — classification scores, risk indices, flagged morphologies — interpreted by a qualified haematologist. Computational hematology is explicitly demarcated from pharmacokinetic modelling (haematological pharmacology), surgical haematology, and non-quantitative pathological description.

Haematology & Pathology

Provides clinical ground truth, WHO 2022 classifications, morphology criteria, and disease ontology.

WHO classification framework
ELN risk stratification
Manual differential methodology

Computer Science & AI/ML

Delivers algorithms, model architectures, training paradigms, and deployment infrastructure.

CNNs, Vision Transformers, MIL
Federated learning (FedAvg)
Foundation models (TITAN)

Laboratory Medicine

Governs instrumentation standards, quality control frameworks, and interoperability protocols.

ICSH / CLSI guidelines
DC impedance, RF, laser scatter
HL7/FHIR · ISO 15189

Bioinformatics & Genomics

Integrates multi-omics data, variant annotations, and population genomic datasets.

scRNA-seq · WGS · NGS
H3Africa genomic datasets
Variant calling pipelines

Section 2

Historical Timeline — Five Eras

The development of computational hematology is understood through five sequential eras, each defined by a paradigm shift in analytical methodology:

Era I · Antiquity — 1880s

Ancient & Pre-Microscopic Observation

Galen’s humoral theory (phlegm, blood, yellow bile, black bile) dominated for 1,500 years. Blood was observed macroscopically — colour, viscosity, clotting — without cellular recognition. Bloodletting (phlebotomy) was the primary intervention. No concept of individual blood cell types existed.

No named inventors — paradigm dominated by Galenic medical tradition

Era II · 1879 — 1950

Manual Microscopy & Staining

Paul Ehrlich described mast cell granulation and introduced acid/base differential staining (1879). Dimitri Romanowsky developed the polychrome methylene blue-eosin stain (1891), refined by James Wright (Wright stain, 1902). Manual leukocyte differential counting on glass slides became standard practice, requiring trained microscopists to classify cells by morphological criteria that persist in WHO classification to this day.

Ehrlich (1879) · Romanowsky (1891) · Wright (1902) · Manual 100–200 cell differential established as standard

Era III · 1950 — 1990

Electronic Counting — The Coulter Principle

Wallace H. Coulter filed U.S. Patent 2,656,508 (1953) for Means for Counting Particles Suspended in a Fluid, establishing the impedance-based cell counting principle that underpins all modern haematology analysers. Mack Fulwyler invented the first cell sorter at Los Alamos (1965). Leonard Herzenberg and colleagues at Stanford patented the first Fluorescence Activated Cell Sorter (FACS) in 1972, with Becton Dickinson commercialising the FACS-1 in 1974. The term “flow cytometry” was formally adopted at the American Engineering Foundation Conference, Pensacola, 1978.

Coulter: US Patent 2,656,508 (1953) · Fulwyler cell sorter (1965) · Herzenberg FACS patent (1972) · BD FACS-1 commercial launch (1974) · VCS technology (Volume/Conductivity/Scatter) introduced 1990s

Era IV · 1995 — 2015

Digital Image Analysis & Classical ML

CellaVision AB introduced the DM96 digital morphology analyser (FDA-cleared; commercially available from c.2001), enabling automated pre-classification of peripheral blood smear cells via artificial neural networks. Whole slide imaging (WSI) platforms emerged for haematopathology. Classical machine learning techniques — support vector machines (SVM), random forests, k-nearest neighbours — were applied to flow cytometry data. Unsupervised clustering algorithms FlowSOM (Van Gassen et al., 2015) and Phenograph (Levine et al., 2015) transformed flow cytometry analysis. Dimensionality reduction via t-SNE (van der Maaten & Hinton, 2008) and UMAP (McInnes et al., 2018) enabled visual cluster exploration of high-dimensional cytometry data.

CellaVision DM96 (2001) · FlowSOM (2015) · Phenograph (2015) · t-SNE (2008) · UMAP (2018)

Era V · 2015 — Present Current

Deep Learning, Foundation Models & African Computational Infrastructure

The publication of Matek et al. (2019) in Nature Machine Intelligence demonstrated human-level blast cell recognition from bone marrow images using ResNet convolutional neural networks — catalysing the deep learning era in haematology. Multiple Instance Learning (SC-MIL, arXiv:2303.13405v2, 2023) addressed the annotation scarcity problem for whole slide images. The TITAN foundation model (2024) enabled zero-shot haematopathology analysis. Federated learning architectures began enabling multi-site training without patient data centralisation. Generative AI (GANs, diffusion models) augmented training datasets. EEHLSS initiated the MedLabAI-LIS platform and BloodBankAI for African clinical deployment.

Matek et al. Nature Machine Intelligence (2019) · SC-MIL arXiv:2303.13405v2 (2023) · TITAN foundation model (2024) · Nazha et al. Blood ASH review (2025) · EEHLSS MedLabAI-LIS (2026)

Section 3

Key Terminology — Three Themed Tables

Table A — Clinical Haematology Terms

Term	Full Form / Context	Clinical Significance & AI Relevance
PBS	Peripheral Blood Smear. Glass slide prepared from EDTA anticoagulated blood; Wright-Giemsa or Leishman stained.	Primary morphological diagnostic tool. WHO 2022 classification criterion for AML, MDS, and lymphoproliferative disorders. ICSH (2015) provides standardised review indications.
CBC	Complete Blood Count. Automated quantitation: WBC, RBC, Hgb, Hct, MCV, MCH, MCHC, RDW, PLT, MPV.	First-tier diagnostic screen. AI models trained on CBC indices achieve >91% accuracy for iron-deficiency anemia differentiation (Kamalzadeh et al., 2025). Foundation of MedLabAI-LIS data capture.
WBC Differential	Manual or automated enumeration of leukocyte subpopulations. CLSI H20-A2 specifies 400-cell manual differential methodology.	Absolute counts preferred over relative percentages (ICSH 2015). Relative neutropenia may be masked by leukocytosis unless absolute counts are reported.
Blast cells	Immature haematopoietic precursors. High nuclear:cytoplasmic ratio; prominent nucleoli; fine, open chromatin.	≥20% blasts in BM or PB = AML diagnosis (WHO 2022). Exceptions include t(8;21), inv(16)/t(16;16), t(15;17) — diagnosed regardless of blast %. AI blast quantification rivals cytologist accuracy (Matek 2019).
Flow Cytometry	FC / Multiparameter Immunophenotyping. Single cells interrogated by laser beam; simultaneous measurement of FSC, SSC, and fluorescent markers.	Gold standard for lineage assignment in haematological malignancy. FlowSOM and ML-automated gating reduce inter-operator variability. Resource-limited adaptation critical for Africa.
Reticulocyte	Immature erythrocyte retaining residual RNA. Detectable by supravital staining (brilliant cresyl blue) or fluorescent flow (CD71/thiazole orange).	Reticulocyte Haemoglobin content (RET-He / CHr) predicts functional iron deficiency. ML models using reticulocyte maturation indices (LFR, MFR, HFR) demonstrate high IDA discrimination.
Bone Marrow Aspirate	BMA. Obtained from posterior superior iliac spine (PSIS) or sternum. Spread on slides; 500-cell differential performed. Perls’ Prussian Blue stain for iron.	Required for MDS staging (ring sideroblasts), AML diagnosis (blast %), myeloma assessment, and aplastic anaemia evaluation. CNN models (Matek 2019) achieve human-level classification on digitised BMA slides.

Table B — Haematological Morphology Terms

Morphological Feature	Cell Type / Location	Clinical Implication & AI Relevance
Auer rods	Needle-like azurophilic cytoplasmic inclusions. Present in myeloid blasts; PBS and/or BMA.	Pathognomonic of myeloid lineage. Diagnostic criterion for AML (WHO 2022). Matek et al. (2019) CNN achieves 94% sensitivity for Auer rod detection. A single Auer rod, in context, is sufficient for AML diagnosis.
Schistocytes	Fragmented erythrocytes. PBS only. ICSH grading: 1+ (<1%), 2+ (1–2%), 3+ (>2%).	≥1% schistocytes = diagnostic threshold for TMA (ICSH 2012 consensus). Key feature of TTP, HUS, DIC, mechanical haemolysis. AI grading eliminates inter-observer coefficient of variation that hampers manual counting.
Smudge cells	Bare lymphocyte nuclei — artefact of smear preparation from fragile lymphocytes. PBS.	>30% smudge cells supports CLL diagnosis; reduced by albumin pre-treatment (Singh & Mori modification). CLL lymphocytes are mechanically fragile due to reduced vimentin expression.
Howell-Jolly bodies	Small, round, dark-staining DNA remnants inside erythrocytes. PBS only; not visible on BMA.	Marker of asplenia (post-splenectomy) or hyposplenism (SCD, coeliac disease, amyloidosis). In SCD, progressive splenic sequestration results in functional asplenia by age 5. AI detects as high-density punctate inclusions within RBC.
Hypersegmented neutrophil	Neutrophil with ≥5 nuclear lobes; ≥1 cell with ≥6 lobes is diagnostic. PBS.	Deficiency of vitamin B12 or folate (megaloblastic haematopoiesis). Also seen post-iron therapy (transient), in iron deficiency, and renal failure. CNN lobe-counting automation reduces variability in megaloblastic anaemia diagnosis.
Pelger-Huët anomaly	Bilobed (“peanut-shaped”) or unilobed neutrophil nucleus with coarse chromatin. PBS.	Inherited (heterozygous) form: harmless. Acquired (pseudo–Pelger-Huët) in MDS: high prognostic significance; indicates dysgranulopoiesis. AI shape descriptor algorithms identify bilobed morphology with high specificity.
Ring sideroblasts	Erythroblasts with ≥5 Perls-stained iron granules encircling ≥1/3 of the nucleus. BMA; Perls’ Prussian Blue stain required.	≥15% ring sideroblasts (or ≥5% with SF3B1 mutation) = MDS-RS subtype definition (WHO 2022). Indicates mitochondrial iron accumulation. Counting is labour-intensive; automated BMA image analysis can quantify sideroblastic ring extent.

Table C — Computational & AI/ML Terms

Term	Definition	Application in Computational Haematology
CNN	Convolutional Neural Network. Deep learning architecture employing convolutional filters to extract spatial features from image data. Subarchitectures: ResNet, VGG, EfficientNet, InceptionV3.	Dominant architecture for blood cell classification from PBS images. ResNet50 and EfficientNet B4 achieve human-level performance on the Munich AML Morphology dataset (Matek 2019). CellaVision pre-classification uses ANN-based CNN.
WSI	Whole Slide Image. Digitised glass slide at 20× (0.5 µm/pixel) or 40× (0.25 µm/pixel) magnification. Gigapixel file format (SVS, NDPI, MRXS).	BMA analysis at slide level; lymph node pathology in lymphoma. Input format for TITAN foundation model. Requires tile-based processing due to gigapixel scale. MIL frameworks address the computational challenge.
Transfer learning	Utilisation of weights pre-trained on a large source dataset (e.g., ImageNet) as initialisation for fine-tuning on a smaller domain-specific dataset.	Overcomes African data scarcity: ResNet50 pre-trained on ImageNet, fine-tuned on ≤500 annotated haematology images. Reduces training time from weeks to hours. Fundamental strategy for resource-constrained deployments.
Grad-CAM	Gradient-weighted Class Activation Mapping. XAI method producing a heatmap that localises image regions most influential for the CNN’s classification decision.	Enables haematologist validation of AI findings (e.g., confirming Auer rod detection is based on the correct cytoplasmic region). Required for regulatory compliance (NAFDAC, SFDA, CE-IVD) and clinical trust.
Data augmentation	Synthetic expansion of the training dataset via geometric (rotation, flipping, elastic deformation) and photometric (colour jitter, brightness, contrast) transformations.	Critical for class-imbalanced haematology datasets where rare cell types (blasts, Auer rods, ring sideroblasts) are underrepresented. GAN-based synthetic PBS cell generation (Shen et al. 2021) extends this approach.
t-SNE / UMAP	Dimensionality reduction algorithms projecting high-dimensional data to 2D/3D for visual cluster analysis. t-SNE: van der Maaten & Hinton (2008); UMAP: McInnes et al. (2018).	FlowSOM + UMAP replaces manual gating in flow cytometry immunophenotyping. Reveals novel cell populations not identifiable by conventional hierarchical gating. Phenograph (Levine 2015) adds community-detection clustering.
MIL	Multiple Instance Learning. Weakly supervised framework learning from bag-level labels without requiring annotations for individual instances within the bag.	WSI-level diagnosis (e.g., slide classified as AML vs. normal) without cell-by-cell annotation — overcoming the annotation bottleneck in African hospitals. SC-MIL (arXiv:2303.13405v2, 2023) demonstrates strong performance on haematology WSIs.
Federated learning	Distributed ML training paradigm: model weights (not patient data) are shared between sites; a central aggregation server applies FedAvg or FedProx to combine updates.	Enables pan-African hematology AI training across LUTH, UNTH, LAUTECH without patient data leaving institutional networks. Preserves data sovereignty. Critical for compliance with Africa CDC Continental Health Data Governance Framework (2025).

Absolute vs. Relative Differential Counts — Clinical Note
A patient with WBC 30.0 × 10⁹/L and 15% lymphocytes has an absolute lymphocyte count of 4.5 × 10⁹/L — within the adult reference interval (1.0–4.8 × 10⁹/L) — whereas the relative value appears elevated. ICSH (2015) mandates reporting both absolute and relative counts. Exclusive reporting of relative percentages may lead to erroneous interpretation of leucocyte subset counts in the presence of leukocytosis or leucopenia.

Section 4

How Modern Analysers Work — 6-Step Pipeline

Contemporary haematology analysers employ multi-modal sensing — direct current (DC) impedance, radio-frequency (RF) conductivity, and laser light scatter (forward scatter FSC, side scatter SSC, axial light loss ALL) — to generate multi-dimensional cellular signatures. The following six-step pipeline describes the flow from sample acquisition to clinical report:

Sample Acquisition & Preparation

EDTA-anticoagulated whole blood is aspirated and diluted. Dedicated lysing reagents selectively destroy erythrocytes and chemically differentiate WBC subpopulations (e.g., exposing nuclear density). Separate dilution channels preserve RBC and platelet morphology.

Multi-Modal Cell Detection

DC impedance (Coulter Principle): measures cell volume. RF conductivity: penetrates cell membrane to assess nuclear:cytoplasmic ratio and nuclear density. FSC: forward light scatter correlates with cell size. SSC: side scatter reflects internal complexity/granularity. ALL: axial light loss measures cellular opacity. Fluorescent channels: detect nucleic acid (reticulocytes, NRBCs) and surface markers.

Multi-Dimensional Clustering

Raw signal vectors (per cell) are submitted to clustering algorithms — Gaussian mixture models, k-means, or support vector classifiers — operating in 3D to 5D feature space. Clusters correspond to: neutrophils, lymphocytes, monocytes, eosinophils, basophils (5-part differential), plus reticulocyte, NRBC, and immature granulocyte sub-clusters.

Automated Flagging & Rule Engine

A deterministic expert rules engine evaluates cluster morphology against pre-defined criteria: blast present flag, immature granulocyte flag, NRBC flag, variant lymphocyte flag, platelet clump flag. ICSH 2015 international consensus criteria define the mandatory PBS review indications when specific flags are triggered. Flagging thresholds must be validated for each analyser-population combination.

Digital Morphology / PBS AI Review (if flagged)

An automated slide maker-stainer (e.g., Sysmex SP-50) prepares and stains the PBS. A digital morphology platform (CellaVision DM96/DM1200 or Morphogo for BMA) captures cell images at the monolayer. A CNN pre-classifies 200–500 cells into morphological categories. The Medical Laboratory Scientist reviews, verifies, or reclassifies results. The final verified differential is released.

LIS Integration & Clinical Decision Support

Results are transmitted via HL7 v2.5 / FHIR R4 to the Laboratory Information System (LIS). Delta check rules compare current results against the patient’s prior values. AI risk stratification modules (as in MedLabAI-LIS) generate supplementary clinical interpretation prompts. The complete report is released to the clinician via the electronic patient record.

Absolute vs. Relative Differential — Analyser Context: Modern analysers report absolute counts directly from cell enumeration; relative percentages are calculated. The absolute count should always be used for clinical decision-making. Analyser-derived 5-part differentials carry an imprecision (CV) of 5–15% at low counts; manual PBS review is mandatory whenever clinical decisions depend on precise WBC subset quantification.

Section 5

2-Year Curriculum Overview — 8 Phases, 72 Weeks

The EEHLSS Computational Hematology Curriculum spans 72 weeks across two years, organised into eight thematic phases. The curriculum is aligned with FRCPath Part I Haematology examination objectives and incorporates Nigeria-specific regulatory context (MLSCN, NAFDAC, NHIA) and GCC deployment considerations (SFDA, CBAHI).

Year 1 — Phases 1–4 (Weeks 1–36)

Phase 1 · Weeks 1–9

Foundations of Computational Haematology

Definitions · history (5 eras) · key terminology · analyzer technology · CBC interpretation · erythrocyte disorder morphology · PBS technique (ICSH 2015) · absolute vs. relative counts

Phase 2 · Weeks 10–18

ML for Blood Cell Morphology

CNN architecture (ResNet, EfficientNet, ViT) · WSI processing (tiling, MIL) · transfer learning · data augmentation (geometric + GAN-based) · leukocyte classification benchmarks · Matek 2019 replication

Phase 3 · Weeks 19–27

Flow Cytometry & Computational Immunophenotyping

FC physics (FSC/SSC/fluorescence) · laser configurations · FlowSOM / Phenograph · t-SNE / UMAP gating · ML immunophenotype classification · automated blast lineage assignment · CD antigen panels

Phase 4 · Weeks 28–36

AI in Haematological Malignancy

AML · MDS · CLL · Lymphoma · WHO 2022 · ELN 2022 risk stratification · blast classification CNNs · Auer rod AI detection · survival ML models (random forest, Cox-LASSO) · MRD monitoring

Year 2 — Phases 5–8 (Weeks 37–72)

Phase 5 · Weeks 37–45

Genomic Haematology & Multi-Omics Integration

scRNA-seq · NGS variant calling pipelines · FISH/cytogenetics AI · multi-omics data fusion · H3Africa genomic datasets · African population genomics · computational CHIP/CCUS detection

Phase 6 · Weeks 46–54

Infrastructure · Federated Learning · LIS Architecture

HL7 FHIR R4 · OpenVINO edge inference · federated aggregation (FedAvg / FedProx) · MedLabAI-LIS data pipeline · MLSCN / NAFDAC AI regulatory frameworks · ISO 15189 AI alignment

Phase 7 · Weeks 55–63

AI Ethics · XAI · Regulation & Governance

Grad-CAM / SHAP / LIME · algorithmic bias and debiasing · NAFDAC / SFDA AI device frameworks · consent-by-design data collection · ISO 15189 AI quality systems · Africa CDC data governance

Phase 8 · Weeks 64–72

Capstone — African Clinical Deployment Projects

SCD / malaria AI pipeline build · BloodBankAI demand forecasting project · federated network implementation · WHX Lagos demo preparation · research manuscript writing · peer-review submission

Section 6

Week 1 Assignment — Lab Challenge Documentation

Each student is required to identify and document five (5) laboratory challenges encountered in their institutional or clinical context, applying structured root cause analysis methodology (5 Whys: Ohno, 1988; or Ishikawa fishbone: 4M/6M framework) and proposing a specific, evidence-based computational solution for each.

Evaluation Rubric

25%

Challenge description & clinical context

30%

Root cause analysis (5 Whys or Ishikawa)

35%

Computational solution proposal

10%

Academic style & citations

Field	Prompt / Instruction	Category Options / Guidance
Challenge title Required	State the specific laboratory problem clearly and concisely using precise haematological terminology. Avoid informal language (e.g., not “smear was bad” → “inadequate PBS preparation resulting in excessive cellular distortion”).	—
Challenge category Required	Select the primary domain of the challenge and justify the selection in ≤2 sentences of passive voice, formal prose.	Morphological · Analytical · Workflow · Data / Informatics · Regulatory · Infrastructure
Clinical context Required	Describe the patient/laboratory scenario. Include relevant CBC parameters, disease context, or institutional setting. Nigerian / West African / GCC contexts are preferred. Reference a specific disease (SCD, malaria, AML) where applicable.	Nigerian tertiary hospital context preferred. SCD / malaria emphasis encouraged.
Root cause analysis High weight	Apply the 5 Whys method (Ohno, 1988) — cascade five iterations from symptom to root cause — or construct an Ishikawa fishbone diagram using the 4M model (Man, Machine, Method, Material). Distinguish proximate cause (immediate trigger) from systemic cause (underlying structural factor).	5 Whys: produce a numbered cascade. Ishikawa: identify ≥3 contributing branches. Both approaches acceptable; justify choice.
Computational solution Primary criterion	Propose a specific AI/ML-based solution. Specify: (1) algorithm or model type, (2) data requirements and source, (3) training strategy (supervised / transfer / federated), (4) validation metrics (AUC, F1, sensitivity/specificity), (5) feasibility evidence from ≥1 peer-reviewed publication.	CNN · SVM · Random Forest · Federated Learning · NLP · Survival model · GAN augmentation · Foundation model (TITAN / SC-MIL)
Implementation barriers Required	Identify ≥2 barriers specific to sub-Saharan African or GCC contexts. Reference regulatory framework or infrastructure limitation where applicable. Propose mitigation strategy for at least one barrier.	Data scarcity · Connectivity · Power infrastructure · Regulatory vacuum · Workforce capacity · Cost
References Min. 2 per challenge	Cite using Vancouver numbered style [1]. Include DOI where available. ≥1 primary research paper required. Textbooks admissible as secondary reference. Maximum 3 references per challenge.	Vancouver (numbered) · DOI required if available · PMC/open-access preferred

Submission specifications: PDF format · ≤500 words per challenge · Total ≤2,500 words excluding references · Passive voice and formal register required · Acronyms defined on first use · Tables and figures encouraged but not required · Submit to EEHLSS MedLabAI-LIS student portal

Section 7