Histopathology 1999, 35, 461–467

A neural network approach to the biopsy diagnosis of early

acute renal transplant rejection
P N Furness, J Levesley,1 Z Luo,1 N Taub,2 J I Kazi,3 W D Bates4 & M L Nicholson5
Departments of Pathology and 5Surgery, Leicester General Hospital; 1Department of Mathematics and Computer Science,
University of Leicester; 2Department of Epidemiology & Public Health, University of Leicester, UK; 3Department of
Pathology, Sindh Institute of Urology and Transplantation, Karachi, Pakistan; and 4Department of Pathology, John Radcliffe
Hospital, Oxford, UK

Date of submission 8 May 1998

Accepted for publication 25 April 1999

Furness P N, Levesley J, Luo Z, Taub N, Kazi J I, Bates W D & Nicholson M L

(1999) Histopathology 35, 461–467
A neural network approach to the biopsy diagnosis of early acute renal transplant rejection

Aims: To develop and test a neural network to assist in
the histological diagnosis of early acute renal allograft
rejection.
Methods and results: We used three sets of biopsies to
train and test the network: 100 ‘routine’ biopsies from
Leicester; 21 selected difficult biopsies which had
already been evaluated by most of the renal transplant
pathologists in the UK, in a study of the Banff
classification of allograft pathology and 25 cases
which had been classified as ‘borderline’ according to
the Banff classification in a review of transplant biopsies
from Oxford. The correct diagnosis for each biopsy was
defined by careful retrospective clinical review. Biopsies
where this review did not provide a clear diagnosis were
excluded. Each biopsy was graded for 12 histological
features and the data was entered into a simple single
layer perception network, designed using the MATLAB
neural network toolbox. Results were compared with
logistic regression using the same data, and with
‘conventional’ histological diagnosis. If the network
was trained only with the 100 ‘routine’ cases, its
Keywords: allograft, Banff, biopsy, neural network, rejection
performance with either of the other sets was poor.
However, if either of the ‘difficult’ sets was added to the
training group, testing with the other ‘difficult’ group
improved dramatically; 19 of the 21 ‘Banff ’ study cases
were diagnosed correctly. This was achieved using
observations made by a trainee pathologist. The result
is better than was achieved by any of the many
experienced pathologists who had previously seen
these biopsies (maximum 18/21 correct), and is
considerably better than that achieved by using logistic
regression with the same data.
Conclusion: A neural network can provide a considerable
improvement in the diagnosis of early acute allograft
rejection, though further development work will be
needed before this becomes a routine diagnostic tool.
The selection of cases used to train the network is
crucial to the quality of its performance. There is scope
to improve the system further by incorporating clinical
information. Other related areas where this approach is
likely to be of value are discussed.

Introduction may be detected, but it is not at all specific; there

Acute renal allograft rejection requires a rapid
increase in immunosuppression if the graft is not to
be lost; but in the early stages of rejection the diagnosis
is often difficult.1 An increase in serum creatinine
Address for correspondence: Dr P N Furness, Department of Pathology,
Leicester General Hospital, Gwendolen Road, Leicester LE5 4PW, UK.
e-mail: pnf1@le.ac.uk
q 1999 Blackwell Science Limited.
are many causes of impaired renal function,
including excessive levels of some immunosuppressive
drugs. Furthermore, in grafts with delayed primary
function due to ischaemic damage, serum creatinine is
inevitably elevated and so provides no indication of
rejection. Alternative approaches to the diagnosis of
rejection, including fine needle aspiration2 and urine
cytology3 have not proved popular, and the main
462 P N Furness et al.
approach remains histological assessment of a needle
biopsy.
Unfortunately the histological changes of acute
rejection develop gradually over hours and days, and
in early cases (which are most amenable to treatment)
the diagnosis can be very difficult. Furthermore, some
histological features of acute rejection can be seen in
protocol biopsies from stable grafts.4
The development of the Banff classification of renal
transplant pathology5 raised hopes that procedures for
the evaluation of renal transplant biopsies would be
harmonized and that the accuracy of diagnosis of acute
rejection would be improved. A recent study of the Banff
scheme showed that the former aim has been achieved,
but the latter has not.6 Twenty-one carefully selected
‘difficult’ transplant biopsies were circulated around the
majority of pathologists who report renal transplant
biopsies in the UK. Using the Banff classification
produced more consistent diagnoses than a conven-
tional approach, but the number of correct diagnoses
(as judged against a retrospective clinical review) was
not improved.
We argued that this disappointing result arose
because in the diagnosis of early acute rejection, the
Banff classification concentrates on just one feature,
‘tubulitis’. Other features which have been informally
considered during a ‘conventional’ evaluation of a
transplant biopsy are ignored.
The human brain is quite good at integrating
disparate piece of information to come to a decision in
an informal way, but it is not consistent. To add more
histological features and ask pathologists to calculate
the probability of rejection in a systematic, reproducible
manner would be impractical. Furthermore, there is
little more than ‘expert opinion’ to indicate what weight
should be attributed to each histological feature.
We proposed that a computer-based system could
allow input of more varied data without losing
reproducibility of assessment. In the context of trans-
plant rejection, this method has the advantage that
subsequent review of the clinical history can, in most
cases, provide an unequivocal diagnosis of rejection or
not rejection. This is invaluable in training and in
testing such a system. We have already shown that this
sort of data integration can be achieved using a
Bayesian belief network, which is a form of inference
network.7 This approach has the disadvantage of
relative inflexibility, as the ‘importance’ attached to
each histological feature has to be calculated and
programmed into the network at the onset. A neural
network raises the possibility of greater flexibility; the
process of ‘training’ a neural network would auto-
matically calculate what ‘weight’ should be allocated to
each histological feature. We therefore sought to
develop a simple neural network which might assist
the decision-making process in the diagnosis of acute
renal transplant rejection.
Materials and methods
H I S T O L O G I C A L F E AT U R E S E X A M I N E D
A standard set of histological features was recorded
for each biopsy by a single observer. The list of
features is given in Table 1. Where a ‘Banff ’ definition
of a feature was available, we used that definition.
For other features, we devised our own grading
system as shown in Table 1, after review of the
range of appearances to be seen in the biopsies to be
studied.
CLINICAL CASES
In a study of this type it is essential to define how the
correct diagnosis is identified. We used a retrospective
review of the clinical casenotes. Rejection was defined as
a rise of serum creatinine of at least 15% in the week
preceding biopsy followed either by a fall to within 5% of
baseline within 7 days of treatment, or by progression to
loss of the graft by rejection. Cases where such changes
could have been due to changes in hydration were
excluded. Non-rejection was defined as a ‘protocol’
biopsy with a change of serum creatinine of less than
5% in the week following biopsy with no change in
immunosuppression, or a rise in creatinine which was
clearly identifiable as a consequence of another condi-
tion, and which responded to treatment of that
condition without increasing immunosuppression.6
TRAINING THE NETWORK
To train the network, we took a sequential series of
transplant biopsies from the files of the pathology
department at Leicester General Hospital. Biopsies from
grafts which had been in situ for 6 months or more were
excluded. The details were passed to a member of the
clinical team (MLN) who excluded cases where a
subsequent review of the clinical notes did not provide
a clear diagnosis or exclusion of acute rejection. In this
way we continued through the files until we had 100
suitable cases; 43 cases of definite rejection and 57 of
definite ‘not rejection’. Sections were then withdrawn
from the files and the severity of all 12 features shown in
Table 1 were recorded for each biopsy by a single
observer (JIK).
q 1999 Blackwell Science Ltd, Histopathology, 35, 461–467.

Table 1. Histological features
Histological feature
Tubulitis
Intimal arteritis
Interstitial infiltrates
Oedema
Interstitial haemorrhage
Acute glomerulitis
Activated lymphocytes
Venulitis
Eosinophils
Plasma cells
Arterial endothelial mononuclear cell adherence
Venous endothelial mononuclear cell adherence
TESTING THE NETWORK
Testing the network with a subset of the original 100
cases did not prove to be taxing; complete discrimina-
tion was possible using only two features, the weight of
the interstitial infiltrates and the extent of the interstitial
oedema. We attributed this to the fact that these biopsies
were unselected, many having histologically obvious
diagnoses. We therefore tested the network with two
other sets of transplant biopsies which lad been selected
to represent difficult diagnostic problems, and which
had previously been subjected to intensive study.
We used 21 selected cases which had previously been
used in the UK study of the Banff classification of
transplant pathology (the ‘Banff ’ cases). These had
initially been selected because they had caused diag-
nostic difficulty for one of the authors (PNF), but a
retrospective clinical review had given a clear diagnosis
in every case; they included 10 cases of rejection and 11
cases which were not rejection. They had all been seen
q 1999 Blackwell Science Ltd, Histopathology, 35, 461–467.
Neural network diagnosis of rejection 463
Scoring system
0–3, as in Banff classification5
0–3, as in Banff classification
Percentage area of biopsy showing obvious interstitial infiltration by
lymphocytes
Percentage area of biopsy showing obvious interstitial oedema
0–3: absent; 1–25%; 26–50%; more than 50% of area
0–3, as in Banff classification
0–3: absent; 1–25% of lymphocytes; 26–50% of lymphocytes;
over 50% of lymphocytes
0–3: absent; scanty lymphocytes, few venules; abundant
lymphocytes, all venules
Number per single high-power field in most heavily infiltrated area
Number per single high-power field in most heavily infiltrated area
Present or absent (cells adherent to luminal surface of endothelium;
c/f intimal arteritis)
Present or absent (cells adherent to luminal surface of endothelium;
c/f venulitis)
by the majority of the renal transplant pathologists of
the UK, so we had a measure of how well-experienced
pathologists might be expected to perform in their
evaluation.6 They had also been reviewed by two
transplant pathologists in the laboratory of the chief
architect of the Banff classification of transplant
pathology.5
The second group of ‘difficult’ cases were selected
from the files of the Department of Cellular Pathology,
John Radcliffe Hospital, Oxford, as part of a major
review of that unit’s transplant biopsies, undertaken by
WDB (the ‘Oxford’ cases). They had been graded as
‘borderline’ according to the Banff classification, but
subsequent clinical review had provided a clear
diagnosis of rejection (19 cases) or not rejection (four
cases).
For all these cases, a single H & E stained section and a
single PAS section were studied by a single observer
(JIK), in the absence of any clinical information beyond
the time after transplantation. For each biopsy a ‘score’
464 P N Furness et al.

was allocated for each of the features listed in Table 1, in

accordance with the definitions provided therein.
It should be recognized that the 100 ‘training’ cases
were an unselected series, but the other two sets were
specifically chosen because they had caused diagnostic
difficulty.

N E T W O R K C O N S T RU C T I O N A N D T R A I N I N G

The network used was a simple single layer perception

network available from the MATLAB neural network
Toolbox. We take a weighted average of the inputs
(Figure 1). Then for some threshold (b), for the
hypothesis that ‘this is acute rejection’ we have:
if P < b then reject the hypothesis
if P > b then accept the hypothesis.
The equation:
X
n
wk Ik ¼ b
K¼1
in Figure 1 is that of a hyperplane in n dimensional
space where the weights wk, k ¼ 1,. . ., n, orientate the
plane. The network learns by finding weights so that the
known answers are obtained on a set of learning data.
In this case acceptance and rejection can be distin-
guished by the plane. With n ¼ 2 this amounts to accept
and reject being separated by a line (e.g. Figure 2).
Furthermore, the certainty with which the hypothesis is
accepted or rejected can be estimated by measuring the
distance of the data point in question from the
separating plane.
In order to provide further network training with
more difficult cases, we subsequently repeated the
exercise after having added the Oxford cases to the
training group; we tested this network with the Banff
cases. We then trained the network with data from the
initial 100 cases and the Banff cases, and tested the
result with the Oxford cases.
Figure 2. An illustration of the n-dimensional hyperplane
constructed by the network. The n observations made for each
biopsy place it in a defined position in n-dimensional space; in this
illustration, for clarity, n ¼ 2. The position of the point on one side of
the hyperplane defines whether the network considers this to be a
case of rejection or not. The distance from the hyperplane provides a
measure of the confidence with which the diagnosis has been made.
It has been suggested that neural networks, especially
relatively simple ones such as this, may provide no
advantage over more conventional statistical analysis.8
We therefore analysed the same data using logistic
regression.9
Results
After training with the 100 ‘training’ cases, we tested
the network with a randomly selected subset of 20 of
those cases. This did not prove to be a taking exercise; all
20 were correctly diagnosed. However, when testing the
network with the more difficult sets, which had been

Table 2. Train with 100 random cases: test with ‘Banff’ set (see
text)

Clinical diagnosis

Rejection Non-rejection Total

Predicted: Rejection 2 0 2

Non-rejection 9 10 19

Figure 1. A simple diagram of the layout of the neural network. The

Total 11 10 21
central equation is explained in the text.

q 1999 Blackwell Science Ltd, Histopathology, 35, 461–467.

 Neural network diagnosis of rejection 465

Table 3. Train with 100 random cases; test with ‘Oxford’ set Table 5. Train with 100 random cases and ‘Oxford’ set; test
(see text) with ‘Banff’ set using logistic regression

Clinical diagnosis Clinical diagnosis

Rejection Non-rejection Total Rejection Non-rejection Total

Predicted: Rejection 0 0 0 Predicted: Rejection 7 8 15

Non-rejection 21 4 25 Non-rejection 4 2 6

Total 21 4 25 Total 11 10 21

selected because the diagnosis or exclusion of rejection
had been difficult, the performance of the network was
very disappointing. With the 21 ‘Banff ’ cases, the
results are shown in Table 2. Only 11 correct diagnoses
were achieved.
With the Oxford set, results are shown in Table 3.
Only 4/25 were correct, which is fewer than might be
expected:)y chance. With both sets the network’s errors
were exclusively under-diagnosis of rejection, as one
might expect as a consequence of training with
‘obvious’ cases of rejection, but testing with less severe
cases. At this stage the network was assigning most
significance to the weight of the interstitial infiltrates.
However, when the network had been trained not
only with the 100 training cases but also with the 25
‘Oxford’ cases, its success rate with the ‘Banff ’ cases
improved to 19/21 (Table 4).
This compares favourably with the highest number of
correct diagnoses provided by any single UK pathologist
(18/21),6 with the number of correct diagnoses
produced when the ‘Banff ’ cases were reviewed in the
laboratory of the main architect of the Banff classifica-
tion (15/21) and very favourably with the mean
number of correct diagnoses provided by all the UK
pathologists (64.8% using a ‘conventional’ diagnostic
approach and 63.3% using the Banff classification).6 It
also compares very favourably with the results of logistic
regression, when applied to the same data (Table 5).
If the network was trained with the 100 training
cases and with the 21 Banff cases, the number of correct
diagnoses achieved when testing with the Oxford group
improved to 24/25 (Table 6).
Again, this compares favourably with the results of
using logistic regression with the same data (Table 7).
After training with the more difficult cases, the
network was using all the histological features except
acute glomerulitis. To our surprise, relatively little
weight was assigned to tubulitis, which is the feature
accorded the most prominence in the Banff classifica-
tion of transplant pathology. This is probably because
the test cases were selected as having caused diagnostic
difficulty; by definition, if tubulitis had permitted a clear
diagnostic distinction, such diagnostic difficulty would
not have arisen.
Discussion
Neural networks have been proposed as useful tools in
decision-making in a variety of medical applications
for a number of years (reviewed in 10). A relatively
simple form of inference network, the Bayesian belief
network, has recently received some prominence in the

Table 4. Train with 100 random cases and ‘Oxford’ set; test Table 6. Train with 100 random cases and ‘Banff’ set; test with
with ‘Banff’ set (see text) ‘Oxford’ set (see text)

Clinical diagnosis Clinical diagnosis

Rejection Non-rejection Total Rejection Non-rejection Total

Predicted: Rejection 9 0 9 Predicted: Rejection 21 1 22

Non-rejection 2 10 12 Non-rejection 0 3 3

Total 11 10 21 Total 21 4 25

q 1999 Blackwell Science Ltd, Histopathology, 35, 461–467.

466 P N Furness et al.
Table 7. Train with 100 random cases and ‘Banff’ set; test with
‘Oxford’ set using logistic regression
Clinical diagnosis
Rejection Non-rejection Total
Predicted: Rejection 16 3 19
Non-rejection 5 1 6
Total 21 4 25
pathological literature.11,12 Nevertheless, computer-
based networks or ‘decision support systems’ have not
found widespread use in routine practice, with the
possible exception of automated cervical cytology
screening systems.13,14 Two possible explanations are
evident. In some applications they produce little or no
improvement beyond conventional ‘expert opinion’ or
more conventional statistical approaches such as logistic
regression.8 In most applications they require careful data
collection and input, which is relatively laborious.
Our results indicate that the first of these objections
need not apply in the context of acute allograft
rejection. The diagnosis of acute rejection is clinically
important. It is usually based mainly on interpretation
of the biopsy appearances; such interpretation is often
difficult, even for experienced practitioners.6 We have
shown that it is possible to use observations made by a
trainee histopathologist to address the diagnosis of
acute rejection in difficult cases, and thus produce a
very high proportion of correct diagnoses. When tested
with the 21 ‘Banff ’ cases, this approach allowed the
trainee pathologist to perform better than any single UK
renal transplant pathologist, better than review by staff
in the laboratory which initiated the Banff classification
of transplant pathology,5 and much better than the
average for the UK’s transplant pathologists. The results
were also better than could be achieved by applying
logistic regression to the same data.
It is notable, however, that the quality of the results is
crucially dependent on the quality of the network
training. In this context it must be admitted that there is
a problem with identifying the ‘gold standard’ diagnosis
against which the results of histological examination
are judged. In the field of transplant rejection, our
approach of retrospective clinical review against preset
standards has been widely used; indeed, our standards
were more stringent than have sometimes been used.
Errors in the ‘agreed diagnosis’ cannot be excluded
completely, but at least we were able to compare the
histological diagnosis with some external standard, rather
than comparing the histological diagnosis from the net-
work with the histological diagnosis of an expert histo-
pathologist, which could become a self-fulfilling prophesy.
If our agreed retrospective diagnoses are accepted,
then training the network with ‘easy’ cases provided an
apparently good performance when tested with the same,
‘easy’ cases; but the network was relying heavily on the
weight of the interstitial lymphocytic infiltrates. Lympho-
cytic infiltration is a prominent feature of acute rejection,
but it is known to be not specific. It is therefore not
surprising that at this stage the network did not provide an
acceptable performance when tested with ‘difficult’ cases,
and gave an under-diagnosis of rejection. It was necessary
first to expose the network to data from comparably
difficult cases before good results were achieved.
In this study there was no interobserver variation, as
all the observations were made by one individual (JIK),
but the sensitivity of the network to the quality of the
training data suggests that interobserver variation
could be a significant problem. It might be necessary
for every pathologist who wishes to use such a system to
train it with their own observations. Differences
between institutions and populations may also have a
major impact. We have some preliminary data which
suggests that even if the observer does not change, a
move from the UK to Pakistan results in a network
becoming unusable until it has been re-trained with
local cases (JIK, manuscript in preparation). Although
laborious, this could have considerable advantages, as
the output would then be related directly and repro-
ducibly to diagnoses made by retrospective clinical
review, and problems of interobserver and interinstitu-
tion variation could be almost eliminated.
The second objection, of excessive labour require-
ments, is more difficult to rebut. Our approach requires
the systematic, separate evaluation of several different
histological features from each biopsy, with subsequent
data entry into a computer. The balance of this
argument has perhaps been modified by the recent
wide acceptance of the Banff classification of transplant
pathology, which requires a similar numeric evaluation
of several features ¹ including tubulitis, which in practice
takes longer to evaluate than any other feature. Other
features (such as eosinophil infiltration, lymphocyte
activation) are under evaluation for incorporation into
the Banff scheme. Some of the features which we used
(such as venous endothelial changes) can probably be
dropped without loss of diagnostic accuracy.
Consequently, the work involved in data acquisition
may be little changed by a neural network approach.
Data entry is a separate problem. At present the
network is set up using the mathematical data
q 1999 Blackwell Science Ltd, Histopathology, 35, 461–467.

manipulation program MATLAB, and the interface is
not sufficiently ‘user friendly’ to be acceptable to
histopathologists; but we hope to be able to produce a
dedicated program which is much easier to use.
We have so far considered only the assessment of
histological features, but a diagnosis of acute rejection is
always made in the context of clinical information.
Neural networks are adept at integrating different types of
observation, so there is no reason why clinical informa-
tion should not be incorporated in the same way. For
example, time since transplantation, rate of change of
serum creatinine, and serum levels of immunosuppressive
drugs could all be incorporated. The results would show
the benefits of reproducibility and objectivity, and would
also provide an evaluation of the significance of each
clinical feature as data accumulates. Remaining within
the field of transplant pathology, the evaluation of chronic
changes would permit a similar approach. Here the
importance of developing ‘surrogate markers’ of chronic
rejection has recently been emphasized,15 and the
‘Chronic Allograft Damage Index’ (CADI) has been
promoted.16 A neural network approach would provide
a more logical way in which the CADI data could be
integrated, giving a more appropriate ‘weighting’ to the
significance of each feature and avoiding the unjustified
assumption that the numeric scoring of each feature
represents a linear scale. A network would also allow
clinical features to be incorporated. The result could be an
approach to predicting graft outcome which is collabora-
tive, rather than the competition which currently exists
between different approaches.
In conclusion, we have demonstrated that neural
network technology can improve dramatically the
accuracy of the histological diagnosis of early acute
renal allograft rejection. The approach has the potential
to remove interobserver variation and to integrate
clinical data. The obstacle to translating these improve-
ments to routine practice is the amount of effort which
will be required by histopathologists, mathematicians
and computer programmers.
Acknowledgements
We are grateful to Dr D R Davies and to Dr M Dunnill
for constructive advice on the manuscript and for
permitting access to the Oxford case files.
q 1999 Blackwell Science Ltd, Histopathology, 35, 461–467.
Neural network diagnosis of rejection 467
References
1. Furness PN. Advances in the diagnosis of renal transplant
rejection. Curr. Diag. Pathol. 1996; 3; 81–90.
2. Hayry P, von Willebrand E. Practical guidelines for fine needle
aspiration biopsy of human renal allografts. Ann. Clin. Res. 1981;
13; 288–284.
3. Kyo M, Miliatsch MJ, Gudat F, Dalquen P, Huser B, Thiel G. Renal
graft rejection or cyclosporin toxicity? Early diagnosis by a
combination of Papanicolaou and immunocytochemical staining
of urinary cytology specimens. Transplant 1992; 5; 71–76.
4. Rush DN, Henry SF, Jeffery JR, Schroeder TJ, Gough J. Histological
findings in early routine biopsies of stable renal allograft
recipients. Transplantation 1994; 57; 208–211.
5. Solez K, Axelsen RA, Benediktsson H et al. International
standardization of criteria for the histologic diagnosis of renal
allograft rejection: the Banff working classification of kidney
transplant pathology. Kidney Int. 1993; 44; 411–422.
6. Furness P, Kirkpatrick U, Taub N, Davies D, Solez K. A UK-wide
trial of the Banff classification of renal transplant pathology in
routine diagnostic practice. Nephrol. Dial. Transplant. 1997; 12;
995–1000.
7. Kazi JI, Furness PN, Nicholson M. Diagnosis of early acute renal
allograft rejection by evaluation of multiple histological
features using a Bayesian belief network. J. Clin. Pathol. 1998;
51; 108–113.
8. Cross SS, Bury JP, Stephenson TJ, Harrison RF. Image analysis of
low magnification images of fine needle aspirates of the breast
produces useful discrimination between benign and malignant
cases. Cytopathology 1997; 8; 265–273.
9. Altman D. Practical Statistics for Medical Research. London:
Chapman & Hall, 1991: 351–358.
10. Cross SS, Harrison RF, Kennedy RL. Introduction to neural
networks. Lancet 1995; 346; 1075–1079.
11. Montironi R, Whimster VYT, Collan Y, Hamilton PW, Thompson
D, Bartels PH. How to develop and use a Bayesian belief network.
J. Clin. Pathol. 1996; 49; 194–201.
12. Hamilton PW, Montironi R, Abmayr W et al. Clinical replications
of Bayesian belief networks in pathology. Pathologica 1995; 87;
237–245.
13. Ryan MR, Stastny JF, Remmers R, Pedigo MA, Cahill LA,
Frable WL. PAPNET directed rescreening of cervicovaginal
smears: a study of 101 cases of atypical squamous cells of
undetermined significance. Am. J. Clin. Pathol. 1996; 105;
711–718.
14. Kok MR, Boon ME. Consequences of neural network technology
for cervical screening: increase in diagnostic consistency and
positive scores. Cancer 1996; 78; 112–117.
15. Hunsicker LG, Bennett LE. Design of trials of methods to reduce
late renal allograft loss: The price of success. Kidney Int. 1995; 48
(Suppl.); S120–S123.
16. Isoniemi H, Taskinen E, Hayry P. Histological chronic allograft
damage index accurately predicts chronic renal allograft rejec-
tion. Transplantation 1994; 58; 1195–1198.