Glioma recognition based on polarization laser-induced breakdown spectroscopy and convolutional neural network

Abstract

Currently, the primary treatment for glioma is surgical resection.However, achieving precise resection during surgical process is challenging due to the difficulty in determining glioma boundary. Inadequate resection increases the risk of recurrence, while excessive resection can result in damage to critical functional areas of brain. Therefore, precise identification of glioma boundary holds significant importance. In this paper, we propose utilizing a two-dimensional polarization laser-induced breakdown spectroscopy (LIBS) combined an adaptive iterative re-weighted penalized least squares-two dimensional-convolutional neural network (airPLS-2D-CNN) model for accurate glioma boundary identification. To validate the superiority of this method, its recognition effect is compared with that of eight classification models, including k-nearest neighbor (kNN), partial least squares discriminant analysis (PLS-DA), linear discriminant analysis (LDA), quantitative discriminant analysis (QDA), support vector machine (SVM), artificial neural network (ANN), adaptive boosting (Adaboost) and random forest (RF), which are built based on LIBS polarization spectra and intensity spectra, respectively. The experimental results demonstrate the airPLS-2D-CNN model established using LIBS polarization spectra has the best performance, achieving an accuracy of 90 %, sensitivity of 90 %, specificity of 90 % in the test set.