HemaCAM, a novel self-learning automated digitalization module,

part of the HaemCell solution.

Sylvie Thoinet1, Céline Darnaud1, Franck Seguy2, Thorsten Zerfass3, Yves Boucaud-Maître1

1Saint Joseph Saint Luc Hospital, Lyon, France2 HORIBA Medical, Montpellier, France, 3 Fraunhofer Institute for Integrated Circuits (IIS), Erlangen, Germany

OBJECTIVES

Traceability of patient records is nowadays a priority in order to ensure quality in laboratories.

The presence of numerical data, graphic, and cytological results within the same screen, optimises and verifies the diagnosis in haematology.

This informed approach is at the core of the "HaemCell“ modular solution from HORIBA Medical. This combines a platform for the haematological analysis (ABX Pentra DX120),

data management (ABX Pentra ML) and a system to acquire and automatically recognize cell images (HemaCAM, by Fraunhofer IIS). We tested the HemaCAM, to evaluate its

performance in normal cell identification. Furthermore, we assessed its capability to correctly discriminate between normal and abnormal cells.

ISLH 2010

Adaptive database capable of evolutive learning

Images validated by laboratory experts were entered into the HemaCAM database. This allowed us to define reference cells (~13000) into the

cell identification mathematical model. We standardized parameters of the slide preparation (staining, counting area), and the optimal number

of cells to be counted according to the reference analyser results. We created a database of seven subpopulations (Enriched sub-populations)

of the 18 available in the HemaCAM: polynuclear neutrophils, eosinophils, and basophils, lymphocytes, monocytes, and additionally nuclear

shadows, and large platelets.

Evaluation

Following the establishment of the enriched sub-populations reference database, samples were selected for the evaluation (300 normal and

100 abnormal), based on the results of the reference analyser in the laboratory. They were then tested on the ABX Pentra DX120 and HemaCAM.

The abnormal ones (with flags or alarms) were additionally analysed with a microscope (200 cells were counted for each smear).

Enriched sub-populations

We calculated the sensitivity (class recall) and the specificity (class precision) of the HemaCAM in recognizing and correctly classifying cells into

the seven enriched WBC subpopulations.

Table N° 1 shows the Class Recall: TP / (TP + FN) -> True Positive Rate (TPR), corresponding to the Sensitivity; the Class Precision:

TP / (TP + FP) -> True Negative Rate (TNR) corresponding to the Specificity (TP: True positive, FN: False negative, FP: False positive)

Percentages are detailed for each category of granulocytes and they were very well preclassified (Fig. N 1). The precision obtained with the

identification of normal samples indicated good results for granulocytes: neutrophils (99.2%), eosinophils (98.7%), and 88.9% for basophils.

The lower performance of the basophils resulted from their low prevalence in the training database (38 cells).Precision was good for lymphocytes

(99.2%). Some large or activated lymphocytes needed to be manually reclassified. It was acceptable for monocytes (86.1%), still this last po-

pulation has to be carefully analyzed since it may contain different types of pathological cells such as myeloid blasts. Large platelets (Fig 2) and

nuclear shadows (Fig 3) were properly recognized, with a precision of 96.0% and 97.5%. Correlations and Bland & Altman plots are showed

in Fig 4 for the 5–differential WBC subpopulations.

LAB WORKFLOW

Fig. 1. Polynuclear eosinophils. Fig. 2.Several images of large

platelets correctly classified.Fig. 3. Several images of nuclei

shadows correctly classified.

Table 1. Class recall (sensitivity)

and class precision (specificity)

obtained for the seven enriched

sub-populations)

Fig. 4. Correlations and Bland-Altman graph for the 5 diff WBC sub-populations

Fig. 5. Platelet aggregates.

Abnormal populations

Immature granulocytes, NRBC, platelet aggregates (Fig 5) were not

correctly preclassified because reference images for those popula-

tions are not yet available in the training database.

Because of high cell morphology variability, blasts (Fig 6-7), lympho-

matous cells, and abnormal immature population (Fig. 8), are mainly

classified as "not identifiable“ but could be easily reclassified ma-

nually into the predefined (Fig 9 ) cell categories.

Fig. 6. Blasts correctly classified Fig. 7. Blast with Auer rods in AML Fig. 8. Pseudo Pelger-Huët

abnormal neutrophil. Fig. 9. List of the predefined 18

subpopulations.

Selections vs alarmsDifferential blood count tested on the haematology analyserABX Pentra DX120, are flagged for shift in the normal distribution of leu-cocytes or for the presence of abnormal cells. This triggers a smear reflex.The ABX SPS Evolution ( Slide Preparation System) spreads and stainsthe smears based on standardized protocol, a key step to ensure thequality and the reproducibility of the smear.

STEP 1 : Loading

The HemaCAM module can load up to 8slides at a time on the stage. Slides are visualizedon the screen and selected for the scan. Thenumber of cells to be counted is predefined(setting menu) but can be modified case by case.

STEP 2 : Slide identification

The positive identification is achieved by anautomatic character recognition system, orwith double manual entry.

STEP 3 : Smear cell density

The scan is launched and HemaCAM selectsthe area of the smear where there is the opti-mal red cell distribution. The validation of thisarea is mandatory before cell image overvie-wing.

STEP 4 : Pre-classification

For this evaluation cells acquired from a slideare automatically classified into the 7 of the 18available subpopulations. Each cell of eachcategory must be visualized before beingallowed to be validated.

STEP 5 : Final User Classification

- Tools to help reclassification:

Each image can be enlarged, measured and

reviewed live with the microscope integrated

in the system in order to examine cellular de-

tails and sub-cellular structures

- Comments & Tags

It is possible to add predefined comments on

RBC, WBC, PLT and free text clinical com-

ments.

Cells can be tagged for further expert examina-

tion (red border) or for specific cell comment

(blue border).

STEP 6 : Validation

Cells can be manually reclassified into the

appropriate subpopulation.

All patient information is recorded and can be

recalled at any time. Date and time, number of

counted cells are available for each folder.

Representative pathological cells can be se-

lected (orange border) to be integrated into

the ABX Pentra ML patient report (V 8).

CONCLUSIONSIn this study, the HemaCAM system highlighted new functionalities compared to those of existing systems for automated cell

image acquisition and identification and showed a satisfying performance to classify the 7 sub-populations tested. The quality

of the optical components, combined with a good depth of field, creates a remarkable final image resolution.

The creation of a new database is now in progress to add more sub-populations (WBC, RBC & PLT lineage ).

Further studies will be necessary to evaluate the evolutive learning capacity of the system.

The software and the environment are very intuitive and user-friendly. This ease of use, the rapidity of the learning process, and

the ability to create specific cases make HemaCAM a promising tool adapted to staff training. A complete integration of the

automatic cell recognition (HemaCAM) and the data management (ABX Pentra ML) module offers a great platform for exploiting

the global haematological data of patients.

HaemCell Solution is a perfect answer to the requirements of security and traceability of the laboratory workflow.

6

5

4

3

2

1