Identification of relevant features in /sup 1/H MR tumour spectra using neural networks

In using neural networks to identify tumour cell types from their /sup 1/H MR spectra, we propose an improvement to the classification performance by identifying those metabolites most significant in differentiating between tumour types. We review two existing approaches for assessing 'relevance' and 'sensitivity' of the network's input units, and introduce a new technique that automatically selects the more dominant inputs. In all three techniques, during training, the network recursively removes the input unit least affecting the error function until the performance drops significantly; the remaining input units are then retained as the most relevant. We test the three procedures on a data set with 180 inputs representing reduced-resolution spectra of six normal and tumour animal tissues. We compare the three methods in terms of the number of remaining units necessary for correct classification and their consistency of using different combinations of train/test sets and initial weights. We conclude that while our proposed algorithm produces a more consistent ranking of features, the number and ranking of significant features varies considerably depending on the algorithm used and the initial conditions and that to obtain the reliable feature subsets combinations of two or more algorithms could be used.