Glioblastoma Classification in Hyperspectral Images by Reflectance Calibration with Normalization Correction and Nonlinear Unmixing

In this work, a new normalized reflectance calibration proposal is presented for hyperspectral (HS) images, and evaluated through a nonlinear unmixing classification method. This evaluation was performed on craniotomy HS images to classify regions affected by grade IV glioblastoma tumor. The classification methodology follows a semi-supervised strategy where some pixels in the HS images were manually labeled by a clinical expert. The nonlinear unmixing of the HS images is carried out by using a multilinear model, and the abundances of the estimated end-members are the distinctive features for classification purposes. The evaluation results show that the proposed calibration decreases the variability of the spectral signatures, increasing the classification accuracy compared to the standard methodology of the state of the art. These results demonstrate that the new formulation allows reflectances calibration without losing characteristic features, which allows better separability among classes than with the standard calibration.

VIVO