Libra framework: Low spectral resolution brain tumor classifier for medical hyperspectral imaging

BACKGROUND AND OBJECTIVE: Gliomas remain one of the most challenging tumors in neurosurgery due to their infiltrative nature and poor prognosis. Machine learning approaches based on hyperspectral imaging have shown potential for assisting intraoperative tumor delineation. However, distinguishing between healthy and tumoral tissues becomes even more difficult when low spectral resolution cameras are employed, which increases the need for specific classification strategies tailored to these devices. This study introduces the Libra framework, a configurable classification framework for processing low spectral resolution hyperspectral images for brain tumor classification, based on ensemble learning and an innovative genetic algorithm data distribution technique applied to the base classifiers. METHODS: The framework was evaluated using two independent intraoperative hyperspectral datasets containing healthy tissue, tumoral tissue, blood vessels, and background classes. Libra incorporates multiple base classifiers whose data distribution is optimized through a genetic algorithm, enabling flexible configurations to adapt to different clinical or computational requirements. Its performance was compared against optimized support vector machines, gradient boosting, adaptive boosting, random forests, and one-dimensional neural networks. The evaluation was carried out on unseen patients, using pixel-level precision and sensitivity as primary metrics. RESULTS: The Libra framework achieved better performance in tumor detection tasks, with a 15% improvement in F1-Score (reaching 46.2%) and a 25%-28% increase in tumoral sensitivity (up to 51.1% depending on the Libra configuration) compared to the optimized support vector machine and gradient boosting models. Precision and specificity also showed consistent improvements across configurations, demonstrating the robustness of the ensemble strategy. These improvements were obtained despite the reduced spectral information and the use of low-complexity algorithms, enabled by the ensemble learning and genetic algorithm data distribution mechanisms integrated into the framework. CONCLUSIONS: The results indicate that Libra provides a notable improvement in identifying tumoral tissue under low spectral resolution scenarios, offering a potential tool to support intraoperative decision-making. The framework's configurability and performance across datasets highlight its potential for future integration into real-time surgical guidance systems, subject to further clinical validation and timing analysis.