Ensemble of Artificial Intelligence Classifiers for In-Vivo Identification of Glioblastoma Tumours Using Hyperspectral Images

TThis study introduces a classification methodology designed for the analysis of in-vivo hyperspectral brain images, aimed at the precise identification of affected regions by glioblastoma tumor. Glioblastoma, known for its aggressive nature and diffuse pattern, poses challenges for accurate intraoperative identification. In fact, current identification procedures rely on intraoperative neuronavigation, which is not accurate, neurosurgeon criteria, which can be subjective, and intraoperative histopathology, which is a time-consuming task. As a result, a two-step approach is proposed to classify hyperspectral in-vivo brain images, which involves non-linear spectral unmixing followed by ensembles of well-know artificial intelligence classification algorithms. Two distinct approaches, namely intra and inter-patient, were utilized for end-member extraction, training, and evaluation of the artificial intelligence classifiers. First, the end-members estimation was approached by a supervised perspective with labeled pixels in the HS images. The corresponding abundances and nonlinear interaction levels obtained from non-linear spectral unmixing were used as features for classification. Our experimental results indicate that the intra-patient approach achieved satisfactory performance using a single classification algorithm. However, in the inter-patient approach, the ensemble technique emerged as the preferred choice by its improved performance. In this way, our results highlight the importance of leveraging ensemble methods when labeled data is limited.