Unifying heterogeneous hyperspectral databases for in vivo human brain cancer classification: Towards robust algorithm development

Abstract

Cancer is one of the leading causes of death worldwide, and early and accurate detection is crucial to improve patient outcomes. Differentiating between healthy and diseased brain tissue during surgery is particularly challenging. Hyperspectral imaging, combined with machine and deep learning algorithms, has shown promise for detecting brain cancer in vivo. The present study is distinguished by an analysis and comparison of the performance of various algorithms, with the objective of evaluating their efficacy in unifying hyperspectral databases obtained from different cameras. These databases include data collected from various hospitals using different hyperspectral instruments, which vary in spectral ranges, spatial and spectral resolution, as well as illumination conditions. The primary aim is to assess the performance of models that respond to the limited availability of in vivo human brain hyperspectral data. The classification of healthy tissue, tumors and blood vessels is achieved through the utilisation of different algorithms in two databases: HELICoiD and SLIMBRAIN. This study evaluated conventional and deep learning methods (KNN, RF, SVM, 1D-DNN, 2D-CNN, Fast 3D-CNN, and a DRNN), and advanced classification frameworks (LIBRA and HELICoiD) using cross-validation on 16 and 26 patients from each database, respectively. Results: For individual datasets,LIBRA achieved the highest sensitivity for tumor classification, with values of 38 %, 72 %, and 80 % on the SLIMBRAIN, HELICoiD (20 bands), and HELICoiD (128 bands) datasets, respectively. The HELICoiD framework yielded the best F1 Scores for tumor tissue, with values of 11 %, 45 %, and 53 % for the same datasets. For the Unified dataset, LIBRA obtained the best results identifying the tumor, with a 40 % of sensitivity and a 30 % of F1 Score.