Machine Learning Performance Trends: A Comparative Study of Independent Hyperspectral Human Brain Cancer Databases

Abstract

Cancer is currently one of the leading causes of death worldwide. Innovative methods that allow early and accurate detection of this disease need to be developed to increase the recovery rate of patients, as accurately distinguishing between diseased and healthy brain tissue for safe surgical resection is a challenge. Research has demonstrated that employing hyperspectral imaging in conjunction with machine and deep learning algorithms can serve as an effective method for detecting in-vivo brain cancer. However, development of real-time solutions is essential for neurosurgeons during surgery, which can be engineered using snapshot cameras at the expense of gathering less spatial and spectral information. Therefore, this work studies the behaviour of different classification algorithms on two independent hyperspectral databases of in-vivo human brains, namely HELICoiD and Slim Brain, gathered with pushbroom and snapshot cameras, respectively. Twenty-six images per database, pre- dominantly affected by glioblastoma, were employed to assess conventional classification algorithms (k-Nearest Neighbors, Random Forest, Support Vector Machine), deep learning methods (1D-Deep Neural Network and 2D-Convolutional Neural Network), and advanced classification frameworks (LIBRA and HELICoiD) through cross-validation. The goal was to distinguish between healthy, tumor, and blood vessel tissues. Results showed that regardless of the database and classifier, the performance of the results improved by reducing and balancing the number of training pixels. Moreover, the test set results highlight the significance of employing advanced machine learning models to improve brain tissue classification, especially on low-resolution hyperspectral data. Notably, one of the best algorithm regarding the mean SEN on the tumor class was LIBRA, which achieved values close to 38%, 72%, and 80% on the Slim Brain, HELICoiD with 20 bands, and HELICoiD with 128 bands datasets, respectively. Then, respect to the mean F1 Score, the HELICoiD framework obtains the best values of tumor tissue around the 11%, 45%, and 53% for the aforementioned datasets, respectively. Furthermore, due to the limited availability of HS data from in-vivo human brain tissue, this study also performed a preliminary evaluation on the combination of such databases with the aim of developing more robust and advanced models. ,

Note:
Funding Declaration: This work has been supported by the Spanish Government and European Union (FEDER funds) in the context of TALENT-HExPERIA (HypErsPEctRal Imaging for Artificial intelligence applications) and TALENT-HIPSTER (HIgh Performance Systems and Technologies for E-health and fish faRming) projects (PID2020-116417RB-C42 AEI/10.13039/501100011033, PID2020-116417RB-C41/AEI/10.13039/501100011033). Additionally, this work was completed while Beatriz Martinez-Vega was beneficiary of a predoctoral grant given by the “Agencia Canaria de Investigación, Innovación y Sociedad de la Información (ACIISI)” of the “Consejería de Economía, Conocimientoy Empleo,” which is part-financed by the European Social Fund (FSE) (POC 2014-2020, Eje 3 Tema Prioritario 74 (85%). Also, Manuel Villa was a beneficiary of the Programa Propio I+D+i 2021 of Universidad Politécnica de Madrid.

Conflict of Interests: None.

Ethical Approval: The study protocol and consent procedures for the HELICoiD Database were approved by the Comité Ético de Investigación Clínica-Comité de Ética en la Investigación (CEIC/CEI) of the University Hospital of Gran Canaria Doctor Negrin (CEIC code: 130069) and the National Research Ethics Service (NRES) Committee South Central - Oxford C for the University Hospital of Southampton. Written informed consent was obtained from all the subjects. The Slim Brain database is a multimodal image databasecreated to identify in-vivo human brain tumors. The guidelines of the Declaration of Helsinki were followed, and the acquisition of HS images was approved by the Research Ethics Committee of Hospital Universitario 12 de Octubre, Madrid, Spain (protocol code 19/158, 28 May 2019).

Keywords: Brain tumor, Hyperspectral Imaging, Machine Learning, deep learning, Neurosurgery, Biomedical Engineering