Meeting Banner
Abstract #2560

Artificial Neural Networks in Radiological Predictive Models

Nikolaos Dikaios1, Taiki Fujiwara2, David Atkinson3, Shonit Punwani2

1Department of Medical Physics and Bioengineering, University College London, London, Greater London, United Kingdom; 2Department of Radiology, University College London Hospital; 3University College London, Centre for Medical Imaging

Predictive models are being increasingly employed in radiology as diagnostic aids for cancer detection. A variety of model types exist. Linear discriminant analysis (LDA) models assume linearity, normality and that the input variables are independent, assumptions which may affect classification accuracy. Neural networks (NN) whilst less intuitive, do not make these assumptions and can detect complex non-linear relationships between the input variables. Both LDA and NN are prone to over-fitting. In this work we compared the performance of multilayer perceptron (MLP) artificial NN and LDA models for prediction of transition zone (TZ) prostate cancer (based on quantitative multi-parametric MRI variables) using a leave-one-out (LOO) and a 2-fold cross validation analysis.

Abstract PDF Poster PDF

Keywords

absence accuracies accuracy accurate achieved advantages affect aids antigen apparent architecture arterial artificial assoc assumptions asymptotic avoid batch better bioengineering biopsy bootstrap bound built cancer certain characteristic class classification coefficient college complex composed computationally conference confidence conjugate cross curve dataset datasets demanding dependent derived descent detect detection diagnostic diffusion discriminant easy eight employed enhanced error estimating evaluated examined exist feed fitting fold formulations forward gland gradient greater hazards hidden hospital illustrates increasingly independent individual input intelligence interpret interval intuitive joint kingdom layer layers linear linearity logistic maintaining make manner maps medical model models multilayer network networks neural normality observations operating outcomes output parametric partitioned patient patients performance physics popular post predicted predicting prediction predictive prone propagation prostate quantitative radiological radiology randomly receiver recorded relationships repetitions replicates rule sample samples selected selection stat suggesting table target template thirty trained training transition types unit upper validation variable variables variety versus volume whereas zone