Aims: To develop and test a neural network: to assist in the histological d
iagnosis of early acute renal allograft rejection.
Methods and results: We used three sets of biopsies to train and test the n
etwork: 100 'routine' biopsies from Leicester; 21 selected difficult biopsi
es which had already been evaluated by most of the renal transplant patholo
gists in the UK, in a study of the Banff classification of allograft pathol
ogy and 25 cases which had been classified as 'borderline' according to the
Banff classification in a review of transplant biopsies from Oxford. The c
orrect diagnosis for each biopsy was defined by careful retrospective clini
cal review. Biopsies where this review did not provide a clear diagnosis we
re excluded. Each biopsy was graded for 12 histological features and the da
ta was entered into a simple single layer perception network, designed usin
g the MATLAB neural network toolbox. Results were compared with logistic re
gression using the same data, and with 'conventional' histological diagnosi
s. If the network was trained only with the 100 'routine' cases, its perfor
mance with either of the other sets was poor. However, if either of the 'di
fficult' sets was added to the training group, testing with the other 'diff
icult' group improved dramatically; 19 of the 21 'Banff' study cases were d
iagnosed correctly. This was achieved using observations made by a trainee
pathologist. The result is better than was achieved by any of the many expe
rienced pathologists who had previously seen these biopsies (maximum 18/21
correct), and is considerably better than that achieved by using logistic r
egression with the same data.
Conclusion: A neural network can provide a considerable improvement in the
diagnosis of early acute allograft rejection, though further development wo
rk will be needed before this becomes a routine diagnostic tool. The select
ion of cases used to train the network is crucial to the quality of its per
formance. There is scope to improve the system further by incorporating cli
nical information. Other related areas where this approach is likely to be
of value are discussed.