A neural network approach to the diagnosis of early acute allograft rejection

Abstract

THE DEVELOPMENT of the Banff classification of renal allograft pathology is intended to improve the reproducibility and the accuracy of the diagnosis of acute renal allograft rejection. These are two separate concepts. A recent trial involving most of the renal transplant pathologists in the United Kingdom found that using the Banff classification improved reproducibility, but accuracy of diagnosis was unchanged from a conventional approach. We argued that this could arise because the Banff classification concentrates on the tight definition of a small number of features (tubulitis, intimal arteritis) but ignores “softer” evidence of transplant rejection such as edema, lymphocyte size, eosinophilic infiltration, and so forth. We have previously reported using a Bayesian Belief inference network to integrate the histopathologic data. When tested using 21 selected difficult transplant biopsies, all of which had clear retrospective clinical diagnoses of acute rejection or not, a trainee pathologist obtained 19 of 21 correct diagnoses. When the same “test” cases were seen by 31 consultant renal transplant pathologists, the best individual performance was 18 of 21. However, the Bayesian network has limitations in its flexibility, which led us to develop a single layer neural network, using the MATLAB neural network toolbox. This network was initially trained with observations from 100 randomly selected renal transplant biopsies, all with clear retrospective diagnoses, using gradings of 12 morphologic features: tubulitis, intimal arteritis, interstitial lymphocytic infiltrates, interstitial edema, interstitial haemorrhage, acute glomeralitis maximal numbers of large “activated” lymphocytes, plasma cells, and eosinophils, venulitis, arterial endothelial activation, and venous activation. When tested using the 21 selected “difficult” biopsies used in the earlier study, the initial performance of the network was disappointing; all but three were graded as negative for rejection. We therefore added a further 25 cases to the training set; these were all selected to have caused diagnostic difficulty at the time of biopsy, but the subsequent clinical course provided a clear diagnosis of rejection or no rejection. After training with this set the network performance improved dramatically, with 19 of 21 correct diagnoses. Conventional logistic regression produced inferior results; only 8 of 21 correct diagnoses, even when including all 125 “training” cases. It appeared that unlike the neural network, logistic regression was “misled” by the inclusion of “obvious” cases in the 100 training set. These results show that logical, reproducible integration of multiple morphologic variables can be achieved by a computer-based neural network approach in a way that can out perform the informal data integration capabilities of the pathologist and that is better than conventional logistic regression. It is likely that the power of this approach can be improved by the inclusion of other types of information, such as clinical and biochemical features. The network currently devised is not in a “user-friendly” format, but there is no technical reason why a program using this approach could not be produced in a way that continues to “learn,” if retrospective validated diagnoses are included. If applied in different centres, such training would automatically make the network adapt to local clinical and pathologic practice, making interobserver variation much less important than at present.