Routine laboratory medicine still depends on trained technologists who look at smears, plates and sediment images. Image-classification models now pre-sort these images: they propose the cell class, flag plates with no growth or suspected pathogens and classify urine particles, so humans validate rather than search. The software ships inside regulated analyser and lab-automation systems (IVDR devices), not as a free-standing chatbot, which makes this a "buy and integrate" use case.
Problem & value
- Today: a manual 100-cell differential on a stained smear takes minutes per slide; culture plates are incubated, carried to a bench and read by eye, often several times; urine sediment is reviewed after analyser flags. Work is repetitive, shift-dependent and short of staff, and inter-reader variability is well documented.
- AI value: the instrument locates and pre-classifies cells, the technologist confirms or reclassifies on screen. Negative plates (often the majority of urine and screening cultures) can be sorted out early, so humans focus on positive or ambiguous plates. Digital images also enable remote review, teaching and second opinions.
- Evidence: peer-reviewed comparisons of the Sysmex/CellaVision DI-60 show good agreement with manual counts for common white-cell classes and improved agreement after technologist verification, with weaker performance on rare or abnormal cells and body fluids [1][2]. Vendor claims on throughput or hands-on-time savings vary by lab and should be re-measured locally.
Variants
- Blood-cell morphology: WBC differential, RBC morphology (anisocytosis, hypochromia, schistocytes), platelet estimate, malaria/parasite and blast screening support; body-fluid cytology as an extension.
- Microbiology plate reading: growth/no-growth triage, colony counting, chromogenic-agar colour classification, presumptive identification, MRSA/VRE screening, antimicrobial-susceptibility-test (AST) inhibition-zone reading.
- Urine sediment: image-based particle classification from flow-imaging or microscopy analysers, combined with dipstick and flow-cytometry flags.
How it works
- specimen + LIS order
- preparation & staining / plating / sediment preparation (lab automation track)
- imaging (smear scanner, plate imager at several incubation times, sediment microscope)
- locate objects (cells, colonies, particles)
- classify (CNN / transformer classifier)
- rules engine (analyser flags, delta check, critical-value rules, age/ward)
- pre-classified result on review screen (technologist confirms / reclassifies)
- autoverification or manual release
- LIS
- clinician report
- feedback of corrections for QC and model monitoring
- Models: convolutional or vision-transformer classifiers trained on very large expert-labelled cell, colony and particle image sets; object detection/segmentation for cell and colony localisation. LLMs and VLMs are at most used for report text or exception explanation, not for the primary classification.
- Human-in-the-loop: a medical laboratory technologist (and for blasts/atypical cells a haematologist or microbiologist) validates every result that is not explicitly allowed to autoverify. Rule-based thresholds decide what is reviewable. Autoverification of negatives needs documented validation and lab-director approval.
Data preparation
Main data are specimen images from slide scanners, plate imagers and sediment microscopes, plus LIS orders and results. They live in the instrument vendor's software, the LIS and local file shares; ground truth sits in technologists' reclassifications and haematology or microbiology reports.
Typical gaps: images in proprietary formats, inconsistent stain or plating protocols between sites, missing or free-text labels, rare classes with too few examples, and patient identifiers burned into images or file names.
Preparation steps for a bought system: confirm instrument and LIS interface (HL7 v2, ASTM, FHIR), clean test and unit catalogues, and assemble a local verification set of 200-500 specimens with expert reference results (PREP-018, PREP-020). For own models: export images in open formats, strip identifiers (PREP-022), have two experts label with a defined class scheme (PREP-019) and capture technologist corrections as feedback (PREP-021).
Tools: vendor export tools, DICOM or standard image converters, a labelling tool such as CVAT or Label Studio, a spreadsheet for agreement statistics. Expect data preparation to be a large share, roughly 40-60 percent, for path C and D, and around 20 percent (verification study) for path A. See GUIDE-003.
Implementation paths
Which path for whom
- enterprise: A - classification of cells, colonies and sediment is a regulated medical function; CE-marked systems with a validated LIS interface avoid taking on manufacturer duties.
- developer: C - own models on open frameworks run on a single GPU (HW1), suitable for research and evaluation, not for clinical use without conformity assessment.
- public: A - university and public laboratories need certified products and tender-compliant procurement; B or C only for research settings.
Path A — Buy (recommended)
- Digital morphology: CellaVision (Sweden, EU; DI-60 co-developed with Sysmex [1], DC-1, CellaVision Vision Hematology software), Sysmex (DI-60), Scopio Labs (full-field digital cytomorphology), Beckman Coulter, Siemens Healthineers and Mindray analysers with image modules.
- Microbiology automation: Copan (Italy, EU; WASP/WASPLab with PhenoMATRIX AI plate-reading software), BD (Kiestra total lab automation with imaging algorithms), bioMerieux (France, EU; automation and ADAGIO/VITEK ecosystem) and Beckman Coulter. Check the CE-IVDR status of each AI module, not only the instrument.
- Urine sediment: Sysmex UF-series and UD-10, Iris/Beckman Coulter and Dirui/Mindray image analysers, Siemens Atellica UAS.
- Choose A because: AI is validated with the instrument, the manufacturer holds the IVDR conformity assessment, and the lab only performs verification, not device development.
Path B — Build on cloud APIs
Rarely applicable for primary reading: general cloud VLMs are not IVDR-validated for this purpose and must not drive diagnostic results. Use EU-region services (Azure, Google Cloud, AWS EU regions, Mistral) and de-identified images; a data-processing agreement is mandatory.
Path C — Self-host (open source)
- Research and internal tooling: PyTorch/MONAI, timm, nnU-Net, Cellpose and StarDist for segmentation, YOLO-class detectors and open foundation models for microscopy; Label Studio or CVAT for annotation; MLflow for tracking.
- Public datasets for prototyping: PBC and Raabin blood-cell sets, the Chula RBC-12 and other published smear sets, DIBaS for bacteria images. Clinical use of a self-built model makes the lab the manufacturer (IVDR Art. 5(5), in-house devices).
- Hardware: HW1 for inference on edge boxes or the instrument PC; HW2 for training small models.
Path D — Custom training
Only for vendors or large academic labs with an in-house-device or manufacturer strategy: needs thousands to hundreds of thousands of expert-labelled images per class across multiple instruments, stain batches and sites; labelling by at least two experts with adjudication; rare classes (blasts, parasites, atypical casts) are the hard part. See DATA-023 for model-in-the-loop labelling.
State of the art (2026)
- Mature: pre-classification of normal WBC classes and common RBC changes is in routine use for over a decade; plate readers reliably separate no-growth plates and colour-classify chromogenic colonies; flow-imaging urine analysers classify common particles automatically.
- Limits and failure modes: rare and abnormal cells (blasts, atypical lymphocytes, parasites) and body fluids are still error-prone and need review [2][3]; stain or slide-quality shifts and new instruments degrade accuracy; polymicrobial plates, mucoid or tiny colonies, swarming and unusual media confuse colony counting; AST zone reading depends on plate quality. Over-reliance (automation bias) lets wrong pre-classifications pass. Performance claims from one population may not transfer.
Shared resources & combinations
- Same imaging stack (image-classification, object-detection, segmentation) as HLTH-002 (digital pathology) and SCI-008 (microscopy); annotation and model-monitoring pipelines are reusable (DATA-023).
- Lab automation: SCI-007 covers robotic lab automation; the track here is the clinical-diagnostic counterpart.
- Upstream/downstream: DOC-007 (lab report extraction), HLTH-028 (LOINC/FHIR mapping for results), HLTH-004 (clinical decision support, e.g. antimicrobial stewardship), PERS-014 (patient-facing explanation of lab results).
Effort, cost & KPIs
- Path A: 3-12 months including instrument procurement, LIS interface (HL7 v2/ASTM/FHIR), verification study and staff training. Cost drivers: instrument lease, reagents/consumables, service contract, LIS integration and validation staff time.
- KPIs: agreement with expert reference (per class, sensitivity/specificity), technologist reclassification rate, turnaround time (specimen to result), share of autoverified negatives, repeat/rerun rate, hands-on time per specimen, missed-pathogen and missed-blast events (zero-tolerance review), EQA results.
People & change
Start with the laboratory director, the responsible haematologist or microbiologist (department-lead) and the medical technologists who will review results daily. Bring in IT for the LIS interface, IT security for instrument networks, the DPO and legal/compliance early, because the system is a high-risk AI system and usually a medical device (MDR, AI Act Annex I route, applicable from 2 Aug 2028 for such products).
Works council: depends. The review screen records which technologist confirmed or reclassified what, so a reclassification-rate dashboard per person would be suited to monitor performance and trigger §87(1) no. 6 BetrVG. If statistics are anonymised at team level the case is weaker, but §90(1) no. 3 information duties remain. Agree in writing that reclassification rates are used only for model QC (GUIDE-004). Patient data: DPIA likely, processing agreement with the vendor (GUIDE-005).
AI literacy: support staff with training on typical failure modes, rare cell types and when to distrust the pre-classification (GUIDE-006). Run a parallel pilot on one bench or shift with a go/no-go on agreement KPIs (GUIDE-009). Typical concerns are deskilling and job loss; address them by keeping technologists as the deciding authority and by limiting autoverification to documented, director-approved cases (GUIDE-008).
Risks, regulation & compliance
- IVDR (EU) 2017/746: AI software that classifies samples for diagnosis is an in vitro diagnostic device; many such products fall into risk class B or C under Annex VIII rules and need notified-body conformity assessment. Transition periods for legacy devices were extended by Regulation (EU) 2024/1860 [4]; verify the current deadline and certificate status of each product.
- EU AI Act (Reg. 2024/1689): a safety component of, or itself, an IVD that requires third-party conformity assessment is high-risk under Art. 6(1) and Annex I. Provider duties (risk management, data governance, logging, human oversight, accuracy documentation) largely apply from 2 Aug 2028 for Annex I products (deferred by the Digital Omnibus on AI, Reg. (EU) 2026/1744); deployers (labs) must ensure oversight by competent staff and AI literacy (Art. 4) [5]. No Art. 5 prohibited practice is involved.
- Accreditation: ISO 15189 (medical laboratories) requires method verification, documented validation of autoverification and competence of reviewers; German labs follow the RiliBAEK (Bundesaerztekammer guideline) for quality assurance.
- GDPR: images and results are health data (Art. 9); smear images with barcodes or labels must be pseudonymised before any secondary use; hospital processing is covered by Art. 9(2)(h) and national law; vendor remote-service access needs a DPA and logging; cloud-hosted vendor analytics must be EU-hosted.
- Clinical risks: missed abnormal cells or pathogens, automation bias, silent performance drift after reagent/stain/instrument changes. Mitigations: documented verification against manual reference, mandatory review rules for flags and abnormal classes, periodic re-verification, drift dashboards, change control, incident reporting under the IVDR vigilance system.
Getting started
- Map volumes and bottlenecks (differentials per day, plates per day, sediment reviews) and your staffing gap.
- Shortlist IVDR-certified systems from your analyser/automation vendors; ask for the certificate, intended purpose, and published verification studies per sample type.
- Plan a method-verification study on your own specimens (include abnormal and rare-class samples) under ISO 15189.
- Define autoverification and mandatory-review rules with the lab director; interface with the LIS.
- Monitor reclassification rate and EQA monthly; keep a change-control log for stains, media and instruments.
Sources
- Sysmex and CellaVision - integrated digital cell morphology system DI-60 - https://www.selectscience.net/article/sysmex-and-cellavision-launch-new-integrated-digital-cell-morphology-system
- Performance evaluation of the digital morphology analyser Sysmex DI-60 for WBC differentials in abnormal samples (PMC) - https://www.ncbi.nlm.nih.gov/pmc/articles/PMC11192923/
- Analytical performance of the Sysmex DI-60 for body fluid cell differential counts (PMC) - https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10374087/
- Regulation (EU) 2024/1860 amending MDR and IVDR transition periods - https://eur-lex.europa.eu/eli/reg/2024/1860/oj
- Regulation (EU) 2024/1689 (AI Act) - https://eur-lex.europa.eu/eli/reg/2024/1689/oj
- Regulation (EU) 2017/746 (IVDR) - https://eur-lex.europa.eu/eli/reg/2017/746/oj
- Red and white blood cell morphology and hands-on time analysis by DI-60 (PMC) - https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9045635/