Multiclass Brain Tumor Classification Using Hyperspectral Imaging and Supervised Machine Learning

Abstract

Hyperspectral Imaging (HSI) can be used as a non invasive medical diagnostic method when used in combination with Machine Learning (ML) algorithms. The significant captured data in HSI can be useful for classifying different types of brain tissues, since they gather reflectance values from different band widths below and beyond the visual spectrum. This allows ML algorithms like Support Vector Machines (SVM) and Random Forest (RF) to classify brain tissues such as tumors. Predicted results can be used to create visualizations and support neurosurgeons before injuring any tissue. This way neurosurgeons can be more precise, reducing any possible damages on healthy tissues. In this work, a proposal for the classification of in-vivo brain hyperspectral images using SVM and RF classifiers is presented. A total of four hyperspectral images from four different patients with glioblastoma grade IV (GBM) brain tumor have been selected to train models and, therefore, classify them. Five different classes have been defined during experiments: healthy tissue, tumor, venous blood vessel, arterial blood vessel and dura mater. Results obtained suggest that SVM usually performs better than RF, generally achieving up to 97% of mean accuracy (ACC). However, RF performance had better results than SVM when classifying images used during training, obtaining almost 100% of mean ACC for all 5 classes described. This study shows the robustness of SVM and the potential of RF for real-time brain cancer detection.