EEHLSS · Computational Hematology Curriculum · Week 3 of 72

WBC Differential Analysis: Manual vs AI-Based Automated Methods

Morphology · Flow Cytometry · CNN Architecture · Accuracy Comparison · Clinical Review Indications

Module 3 of 8FRCPath Part I
ICSH 2015 · WHO 2022MedLabAI-LISeehlss.io

Section 1

Introduction

The white blood cell (WBC) differential is one of the most clinically consequential haematological tests in routine laboratory medicine, providing both quantitative (absolute and relative counts) and qualitative (morphological) information about the five major leukocyte lineages.

Domain	Haematological Pattern	Clinical Context
Bacterial infection	Neutrophilia + left shift (band forms)	Sepsis, pneumonia, abscess
Viral infection	Lymphocytosis + atypical lymphocytes	EBV, CMV, HIV, COVID-19
Parasitic/allergic	Eosinophilia	Helminths, atopy, drug reaction
Haematological malignancy	Blasts, leukaemic picture	AML, ALL, CLL, MDS, lymphoma
BM function monitoring	Neutropenia, engraftment patterns	Post-chemotherapy, post-transplant

ICSH 2015 mandate: Always report both relative (%) AND absolute counts (×10⁹/L). Relative percentages alone can be profoundly misleading in the context of leukocytosis or leucopenia.

Section 2

Five Leukocyte Lineages

50–70%

Neutrophil

2–5 lobes · fine pink granules · phagocytosis · 12–18 µm · first-line bacterial/fungal defence

20–40%

Lymphocyte

Round dense nucleus · scant basophilic cytoplasm · T/B/NK · 7–12 µm · adaptive immunity

2–8%

Monocyte

Kidney-shaped nucleus · ground-glass cytoplasm · 14–21 µm · chronic infection · → macrophage

1–4%

Eosinophil

Bi-lobed nucleus · large orange-red granules · 12–17 µm · allergy · parasites · IgE

0–1%

Basophil

S-shaped nucleus · dark purple-black granules · 10–16 µm · IgE-mediated · CML · heparin

7-part differential — the left shift

Left shift (%) = (Band neutrophils / Total neutrophils) × 100

> 6% bands OR absolute band count > 700/µL = significant left shift
→ Suggests: acute bacterial infection, severe physiological stress, bone marrow stimulation

Key morphological abnormalities and AI detection rates

Feature	Clinical Association	AI (CNN) Detection
Auer rods	Pathognomonic AML (WHO 2022 ≥20% blast threshold)	88–94% sensitivity
Toxic granulation	Severe bacterial sepsis	Moderate recognition
Döhle bodies	Bacterial infection, pregnancy, burns	Moderate recognition
Hypersegmented neutrophils	B12/folate deficiency, megaloblastic anaemia	CNN lobe-counting feasible
Smudge cells	CLL (fragile lymphocyte membranes)	Recognised (CellaVision)
Atypical lymphocytes	EBV, CMV, HIV — viral reactive	85–92% sensitivity
Blasts	Acute leukaemia (AML/ALL)	88–95% sensitivity
Pseudo-Pelger-Huët	MDS dysgranulopoiesis	AI shape-descriptor approach

Section 3

Manual Differential Methodology

Sample preparation and technique

EDTA-anticoagulated blood analysed within 6 hours of collection (morphological degradation accelerates beyond this). Well-made PBS with feathered edge, ~1 WBC per HPF at 100× oil immersion. Wright-Giemsa staining with pH-controlled buffer — incorrect pH directly alters granule coloration and causes systematic misclassification. Stain pH must be quality-controlled daily.

Systematic traversal using zigzag or perimeter (battlement) method. Minimum 200 leukocytes counted (100 for urgent/STAT). CLSI H20-A2 specifies 400-cell differential as the gold-standard reference method. Oil immersion objective (100×) for definitive cell identification.

Statistical limitation:
95% CI at 50% prevalence (n=200): ±7%
95% CI at 5% prevalence (n=200): ±3%
This sampling imprecision affects sensitivity for rare cell detection

Advantage	Limitation
Detects abnormal morphology: blasts, Auer rods, toxic granulation	Subjective — inter-observer variability especially eosinophil/basophil
Identifies rare populations not captured by automated algorithms	Time-intensive: 5–10 minutes per slide
Simultaneous RBC and platelet morphology assessment	Fatigue-related errors increase in high-volume shifts
Gold standard for discrepant or flagged automated results	Sampling error from uneven smear cell distribution
Qualitative descriptive narrative capability	Statistical imprecision at low-prevalence cell types

Section 4

Automated Technology Platforms

3-part vs 5-part differential: Basic impedance-only analysers produce only a 3-part differential (neutrophils / lymphocytes / “middle cells”). Clinical haematology requires 5-part capability at minimum. The Coulter Principle (US Patent 2,656,508, 1953): DC resistance pulse ∝ cell volume as cells pass through an aperture between two electrodes.

Sysmex XN Series — SFL fluorescence

FSC → cell size (forward scatter)
SSC → granularity/internal complexity
SFL → RNA/DNA (fluorescent nucleic acid stain)
RF → impedance volume
Neutrophils: low SFL + medium SSC
Eosinophils: high SSC (specific granules)
Basophils: high SFL (heparin content)
Differential lysing reagents

Abbott Alinity h — MPO fluorescence

FSC low-angle → cell diameter
SSC 90° → internal granularity
SSC low-angle → nuclear complexity
Yellow fluorescence: nucleic acid
Red fluorescence: myeloperoxidase (MPO)
Neutrophils: high MPO + moderate DNA
Lymphocytes: low MPO + low-moderate DNA
Monocytes: moderate MPO + high DNA

Limitation common to all platforms: Cannot classify blasts or dysplastic cells reliably — flags trigger mandatory manual PBS review. Interference from NRBCs, giant platelets, and cold agglutinins produces spurious results. Limited maturity assessment beyond basic segmentation.

Section 5

CNN Architecture for WBC Classification

Input (high-resolution PBS image, single-cell crop)
→ Convolutional layers: detect edges, textures, nuclear shapes, granule patterns
→ Pooling layers: dimensionality reduction, spatial invariance
→ Fully connected layers: feature combination, cell-type decision (ReLU activation)
→ Softmax output: probability per class (sum = 1.0)
→ Classification: Neutrophil / Lymphocyte / Monocyte / Eosinophil / Basophil / Blast

Feature hierarchy

CNN Depth	Features Extracted	Haematological Correlate
Early layers	Nuclear contour, cell boundary, basic size	Cell size classification; NRBC identification
Middle layers	Granularity, N:C ratio, chromatin texture	Neutrophil vs monocyte; granule density
Late layers	Cell-type combination signatures	Final classification output; rare cell detection

Training strategies

Strategy	Purpose	Implementation
Supervised learning	Expert-annotated ground truth labels	Munich AML (Matek 2019); Acevedo WBC-17
Data augmentation	Overcome class imbalance (rare cells)	Rotation, flipping, colour jitter, noise
Transfer learning	Overcome data scarcity	ResNet50, VGG16, InceptionV3 from ImageNet
Dropout regularisation	Prevent overfitting during training	Training only; removed at inference
Ensemble methods	Reduce single-model variance	Multiple CNN architectures averaged

Performance benchmarks (meta-analysis 2020–2023)

Task	Accuracy Range	Clinical Context
Multi-class WBC classification	92–98%	Clean, well-stained smears
Band neutrophil sub-classification	85–92%	Most challenging sub-class distinction
Blast / atypical lymphocyte detection	88–95% sensitivity	Critical for malignancy screening
vs. junior haematologists	Non-inferior	Benchmark comparison
vs. senior expert morphologists	Approaching	Gap narrowing with larger datasets

CNN critical limitations: Highly sensitive to smear quality and staining consistency — primary performance degradation factor. Out-of-distribution failure on unusual morphologies not in training set. Black-box opacity (Grad-CAM/SHAP required for XAI transparency). FDA/CE-IVD approval required before clinical deployment. LIS integration requires compatibility validation and version control.

Section 6

Accuracy Comparison: Human Morphologist vs AI

Cell Type	Human (%)	AI/CNN (%)	Clinical Interpretation
Neutrophils	94	91	Human advantage — contextual nuclear lobe assessment
Lymphocytes	92	93	AI slight advantage — consistent objective gating
Monocytes	85	88	AI advantage — nuclear folding variability managed objectively
Eosinophils	80	85	AI advantage — granule colour objectivity eliminates observer variance
Basophils	70	75	Both struggle — rarest cell; minimal training/observer exposure
Overall agreement	85–90% between expert morphologists and AI systems

Human advantage

Rare/novel morphology; contextual interpretation with clinical history; simultaneous RBC/platelet assessment; qualitative narrative comments

AI advantage

Consistent objective criteria; immune to fatigue and cognitive bias; high precision for well-defined categories; quantitative trending; rapid throughput

AI disadvantage

Training dataset geographic/ethnic bias; artefact misclassification; cannot contextualise with patient history; OOD failure; black-box opacity

Section 7

Clinical Scenarios Requiring Manual Review

Mandatory manual PBS review: These clinical scenarios cannot be managed by automated differential alone. Manual review by a qualified MLS or haematopathologist is required — ICSH 2015 consensus criteria define specific flag combinations.

Clinical Category	Specific Indication	Key Morphological Target
Haematological malignancy	Acute leukaemia (ALL/AML)	Blast identification + classification; Auer rods (AML)
	Chronic leukaemias (CLL/CML)	Atypical lymphocytes, smudge cells, basophilia
	Myelodysplastic syndrome	Hypogranulation, pseudo-Pelger-Huët, dysplasia
	Lymphoproliferative disorders	Atypical lymphocyte morphology assessment
Infectious disease	Suspected parasitaemia	Malaria ring trophozoites, Babesia, Trypanosoma
	Severe bacterial sepsis	Toxic granulation, Döhle bodies, vacuolisation
	Viral infection (EBV, CMV, HIV)	Atypical reactive lymphocytes
Special populations	Paediatric samples	Different reference ranges; physiologically higher lymphocytes
	Pregnancy	Physiological leukocytosis with left shift
	Geriatric	Increased MDS risk; dysplastic changes
	Post-transplant	Engraftment pattern; GVHD monitoring

Turnaround time requirements

Priority	Requirement
STAT / urgent	Manual review within 30–60 minutes of flag generation
Routine	Batch review within 4-hour window
Critical result	Immediate notification regardless of differential completion status

Section 8

Quality Control and Standardisation

Phase	Variable	Requirement
Pre-analytical	Anticoagulant	EDTA preferred — preserves morphology
	Timing	Analysis within 6 hours of collection
	Temperature	Room temperature — avoid cold agglutination
	Mixing	Proper inversion — prevents leukocyte stratification
Manual differential QC	Stain QC	Daily Wright-Giemsa pH and filtration check
	Competency	Quarterly technologist assessment
	Inter-lab comparison	External QA programme participation
	Classification criteria	ICSH 2015 standardised morphological definitions
Automated platform QC	Daily QC	Tri-level controls (normal, low, high)
	Weekly studies	Linearity and carryover testing
	Monthly extended QC	Abnormal sample panel
	Software control	Version control + clinical validation before update

Section 9

Future Directions

Hybrid human-AI systems

The emerging clinical consensus positions AI as a first-reader with technologist verification of outlier cases, rather than a replacement for expert morphology. Continuous learning systems incorporating expert corrections from clinical feedback loops progressively improve local population performance. Explainable AI (XAI) approaches using Grad-CAM and SHAP provide visual explanation of CNN classification decisions — essential for clinical trust and regulatory compliance.

African-specific deployment considerations (MedLabAI-LIS)

Standard CNN models trained on European and North American datasets show reduced performance on African blood smears where malaria-SCD co-presentation, tropical eosinophilia, and nutritional anaemia produce morphological patterns not represented in training data. EEHLSS MedLabAI-LIS incorporates African population-specific fine-tuning using LoRA rank 16 on LUTH/UNTH institutional datasets with DinoBloom-L as the primary hematology vision encoder (OpenVINO INT8, ~8ms/cell inference on GMKTech local server). Population-specific reference intervals are built into the FHIR R4 DiagnosticReport output.

Advanced technologies

Digital whole-slide imaging for remote haematopathology consultation. Spectral cytometry for enhanced fluorescent parameter measurement. Raman spectroscopy for label-free biochemical characterisation. Point-of-care AI (smartphone microscopy + CNN) for resource-limited and LMIC settings — directly applicable to district hospital-level haematology in Sub-Saharan Africa and the GCC.

Section 10

Assignment & References

Week 3 Assignment

Task (eehlss.io student portal):

1. Review three anonymised blood smear image sets via MedLabAI-LIS. For each: perform a 100-cell manual differential, compare against AI-generated differential output
2. Calculate the left shift percentage for each case. Interpret with reference to ICSH 2015 criteria
3. For the blast-flagged case: describe distinguishing morphological features, state WHO 2022 threshold, propose next three investigative steps
4. For the case where AI and human differentials diverge by >10% in any cell category: identify the likely source of discrepancy and propose a resolution strategy

300 words per case · Minimum 2 Vancouver-style references per case · PDF submission via MedLabAI-LIS portal

References

#	Type	Citation & DOI
1	Textbook	Bain BJ. Blood Cells: A Practical Guide. 6th ed. Wiley-Blackwell; 2015. ISBN: 9781118408889
2	Primary	Matek C, et al. Human-level recognition of blast cells in bone marrow. Nature Machine Intelligence. 2019;1(11):538–544. DOI: 10.1038/s42256-019-0101-9
3	Primary	Acevedo A, et al. A dataset of microscopic peripheral blood cell images. Data in Brief. 2020;30:105474. DOI: 10.1016/j.dib.2020.105474
4	Review	Nazha A, et al. Artificial intelligence in hematology. Blood. 2025;146(19):2283–2292. DOI: 10.1182/blood.2025029876
5	Guideline	ICSH Working Group. Reference method for WBC differential counting. Int J Lab Hematol. 2015;37(3):287–303. DOI: 10.1111/ijlh.12228
6	Textbook	Lee GR, et al. Wintrobe’s Clinical Hematology. 13th ed. Wolters Kluwer; 2018. ISBN: 9781451172683
7	Africa	Maruta T. AI in diagnostics: laboratory medicine in Africa. Afr J Lab Med. 2025;14(1), a2952. DOI: 10.4102/ajlm.v14i1.2952

WK 3 — WBC Differential Analysis