![]()
11
| 4 | 2006
Klinisk Biokemi i Norden
(Fortsætter side 12)
and scanned automatically. Blood cells are local-
ized and images are collected and processed. The
software uses artificial neural network technology.
More than 100 features are calculated from each
cell image; the features are then analysed and used
for the pre-classification of the cells. Following a
number of iterations the system learns to identify
the cells. The training set contains some 35,000
cells classified by an expert panel.
The results are presented on a computer screen,
and all cells of the same class can be studied at the
same time (figure 2). Different classes can be dis-
played simultaneously, or an individual cell can be
zoomed in and studied more closely (figure 3). The
operator always manually verifies the classifications
suggested by the system. If the operator is uncertain
about how to classify cells, a reference library con-
taining pre-classified cells can be displayed on the
screen. This facilitates classification for the operator
who improves and stabilizes the classification of the
cells. Using the same defined set of reference cells
also greatly serves to make the leukocyte classifica-
tion more uniform both within the laboratory and
also between different laboratories.
There are three stages in the process: the first is
cell location, the second is cell pre-classification
and the third is manual validation. For 10-20%
of the cells, reclassification by the technologist is
necessary. In general the performance of the com-
puterised classification declines with increasing
abnormality/immaturity of the cells.
Our evaluation (7) has been performed on 322
routine specimens; half of them normal, half of
Cell class
Neutro-
phils
Eosino-
phils
Baso-
phils
Mono-
cytes
Lympho-
cytes
NRBC Blast
cells
Meta-
myelo-
cytes
Unident-
ified
Smudge
cells
R-value
(corrected/
uncorrected
classification)
0.99
0.96
0.90
0,72
0.98
0.04
0,70
0,33
0,62
0,78
Table 1
Correlation between computerised classification and results obtained after human review of these classifications (N = 44).
Figure 2: A view of the computer screen displaying result of a computerised localisa-
tion and classification of peripheral cells.
Figure 3: A view of a
zoom in on a peripheral
white blood cell.
