Glioblastoma Detection with Hyperspectral Image Analysis through Optimal Wavelength Selection

Abstract

Glioblastoma is the most aggressive and common type of malignant primary brain tumor. Neurosurgery is one of the main treatments for the removal of glioblastoma tumors. Although complete tumor resection is crucial, excessive removal of brain tissue can cause unwanted impairment. Intraoperative techniques for tumor detection and delineation can help to achieve a more precise resection and improve the clinical workflow and outcomes. This study explores the use of hyperspectral imaging for detecting glioblastoma during surgery. To this end, a database of 24 images from 14 patients is studied by employing an image analysis framework, which entails spectral and spatial dimensionality reduction and classification. Multiple AI-based methods are presented and tested for the detection of healthy tissue and glioblastoma, as well as techniques for reducing HSI dimensionality, thereby facilitating the clinical applicability of HSI. A multi-layer perceptron shows the highest macro F1 score of 86.65%, when 20 hyperspectral wavelengths are automatically selected by using the Ant Colony optimizer. The proposed approach outperforms the state-of-the-art methods, which use datasets including multiple grades and solely grade 4 tumors. The results demonstrate that HSI combined with a proper image analysis framework, aiming at reducing spectral and spatial dimension, has the potential to aid tumor detection during brain surgery.