Comparative Evaluation of Feedforward and Probabilistic Neural Networks for the Automatic Classification of Brain Tumours

Abstract

Brain tumours grading is a crucial step for determining treatment planning and patient management. The grade of a tumour is defined by pathologists after reviewing biopsies under the microscope, a procedure that has been proven highly subjective. In this work, we propose a computer-based system for the automatic classification of astrocytomas that can be used as a second opinion tool for the clinicians contributing to the objectification of the diagnostic process. The system process routine brain tumours biopsies and performs automatic diagnosis of the degree of tumour abnormality (low from high grade tumours) based solely on quantitative information acquired from cell nuclei. We designed the system incorporating two stat-of-art neural network algorithms, namely Perry’s nonmonotone spectral conjugate gradient training algorithm for Multi Layer Perceptrons (MLPs) and Probabilistic NN (PNN). Best performance was obtained using a MLP-NN classifier with 7-hidden neurons topology that discriminating low from high grade tumours with an accuracy of 92.0%. Sensitivity and specificity ranged to 93.1% and 90.5% respectively. The PNN classifier resulted in lower rates (83.3% specificity, 91.5% sensitivity and 89.9% overall accuracy). The proposed method is a dynamic new alternative to brain tumour grading since it combines relatively high accuracy rates with daily clinical standards.

6 Figures and Tables

Cite this paper

@inproceedings{Kostopoulos2004ComparativeEO, title={Comparative Evaluation of Feedforward and Probabilistic Neural Networks for the Automatic Classification of Brain Tumours}, author={A. E. Kostopoulos and Dimitris Glotsos and Panagiota Spyridonos and George Nikiforidis and D. G. Sotiropoulos and Theodoula N. Grapsa}, year={2004} }