A pattern recognition system for brain tumour grade prediction based on histopathological material and features extracted at different optical magnifications

Abstract

The purpose of this study is to develop a computer-assisted diagnosis system for improving diagnostic accuracy in brain cancer classification into grades of malignancy. The clinical material comprised biopsies of patients with confirmed brain cancer. Images were digitized from the original material using a digital light microscopy imaging system (LEICA Axiostar plus coupled with a LEICA DFC 420C camera, Leica Microsystems GmbH). The digitized images were processed for the separation of nuclei from the surrounding tissue using edge detection techniques. Then, features were extracted from segmented nuclei at different optical magnifications to describe each sample-patient malignancy status. Moreover, samples were examined by an expert pathologist (P.R.), who assessed qualitative a number of crucial histological characteristics that are used by the World Health Organization as criteria for tumours’ grading. These features comprised the input to a pattern recognition system, which was designed in order to predict the risks of malignancy of each tumor. The system was structured using the Probabilistic Neural Network (PNN) and Support Vector Machine (SVM) classifier alternatively. Using the leave-one-out method, the PNN resulted in 94.4% accuracy, while the SVM showed 96.3%. To assess the generalization of the system to unknown data, the external cross validation was used and gave 77.8% prediction for both classifiers. Results show that computer-assisted diagnosis offers a valuable tool providing second opinion consultancy to expert physicians, which contributes towards a better and more accurate diagnostic conclusion