Brain tumors represent a significant global health concern, with increasing incidence and mortality rates. Surgical resection is a primary treatment for many brain tumors, but accurate delineation of tumor boundaries during surgery is crucial to maximize resection while minimizing damage to healthy tissue. Current intraoperative imaging techniques, such as image-guided stereotactic neuronavigation, intraoperative MRI, and ultrasound, have limitations in terms of accuracy, cost, or invasiveness. Fluorescence imaging, while offering real-time visualization, is limited in its ability to detect low-grade gliomas and requires the administration of contrast agents. Hyperspectral imaging (HSI) emerges as a promising alternative, providing label-free, non-contact, near real-time, and minimally invasive intraoperative guidance. HSI captures hundreds of narrow spectral channels, allowing for the identification of tissues based on their chemical composition. Recent advances in AI and increased computational power have enabled HSI to achieve promising results in various medical applications, including oncology. This study aims to establish a benchmark for intraoperative brain tumor detection using HSI and machine learning, building upon previous work by the research group which utilized a smaller dataset and different validation techniques. The goal is to develop a real-time decision support tool for neurosurgical workflows that can improve patient outcomes.

The literature extensively documents the challenges in intraoperative brain tumor detection and the limitations of existing techniques. Studies highlight the importance of maximal safe resection to improve survival rates for various glioma types. The brain shift phenomenon, cost constraints of intraoperative MRI, and artifacts associated with ultrasound limit the effectiveness of current methods. Fluorescence imaging, while useful for high-grade gliomas, struggles with low-grade tumors. Several research groups have explored the potential of HSI in various medical fields. Studies show its effectiveness in gastrointestinal, head and neck, and skin cancer detection, showcasing improved accuracy compared to standard RGB imaging. However, large, high-quality datasets for HSI-based AI algorithms are often lacking. The research team's prior work demonstrated the proof of concept for using HSI and machine learning for brain tumor detection, but this study aims to build upon those results with a more extensive dataset and a more robust validation approach, providing a stronger benchmark for future research.

This study utilized a custom-designed intraoperative hyperspectral imaging (HSI) acquisition system to collect data from 34 adult patients undergoing brain surgery. The system consisted of a VNIR HS pushbroom camera capturing 826 spectral channels (400-1000 nm). Data collection involved three campaigns, leading to a total of 61 HS images. The images were pre-processed to correct for illumination and dark current, reduce noise using a moving average filter, and remove less informative spectral channels. Dimensionality reduction was achieved by decimating the spectral channels to 128. Ground truth (GT) maps were created by manually labeling pixels into four classes: tumor tissue (TT), normal tissue (NT), blood vessels (BV), and background (BG). Eight different classification algorithms were evaluated: random forest (RF), k-nearest neighbors (KNN) using Euclidean and Cosine distances, support vector machines (SVMs) with linear and RBF kernels, a two-layer deep neural network (DNN), and two unmixing-based methods (linear and nonlinear extended blind end-member and abundance extraction - EBEAE and NEBEAE). To manage computational costs associated with algorithm training, the training dataset was reduced using K-Means clustering to select the most representative pixels. A three-way data partition (training, validation, and test sets) along with 5-fold cross-validation ensured robust evaluation. The performance was assessed using macro F1-score, overall accuracy, sensitivity, and specificity. A spatial-spectral approach was also implemented, incorporating spatial information through KNN filtering and unsupervised hierarchical k-means (HKM) segmentation, which was further combined with a majority voting (MV) approach to generate three maximum density (TMD) maps. Finally, a preliminary post-hoc interpretability analysis was performed using LIME to determine the most relevant wavelengths for classification.

Spectral characterization revealed statistical differences between tumor tissue and normal tissue/blood vessels across various spectral ranges, particularly around HbO2 and deoxyHb absorbance peaks. Higher absorbance values were observed in tumor tissue compared to normal tissue but lower than in blood vessels. Statistical differences were observed among primary and secondary tumors and across different tumor grades (G1-G4). The spectral-based classification (using the validation set) showed that SVM-based and DNN methods yielded the best macro F1-Score results. The inclusion of spatial information improved the macro F1-score medians, although no statistically significant differences were found compared to spectral-only methods. The best median macro F1-score achieved was 70.2 ± 7.9% on the test set using the DNN algorithm with the spatial-spectral approach. Qualitative results demonstrated the system's capability to identify different tumor types (high-grade, low-grade, and secondary). The post-hoc interpretability analysis using LIME indicated that several wavelengths, particularly those associated with HbO2 and deoxyHb absorbance peaks, were consistently identified as important features across multiple classification models. Comparison with previous studies showed that the current study's methodology, employing a larger dataset and a more rigorous validation approach, provided more robust and generalizable results.

The findings of this study demonstrate the significant potential of HSI as a real-time decision support tool for intraoperative brain tumor detection. The high macro F1-score achieved, along with qualitative results, validate the system's ability to identify and delineate various tumor types. The improved performance compared to previous studies underscores the importance of using larger datasets and robust validation methods. While some limitations exist regarding the challenges of optimal image acquisition in certain conditions (e.g., deep-layer tumors), the DNN model showed a capability to overcome some of these limitations. The spatial-spectral approach further improved the classification accuracy, showcasing the synergistic benefits of combining spectral and spatial information. The post-hoc interpretability analysis sheds light on the spectral features that are critical for accurate tumor identification, providing insights for further model improvement. Future work should focus on addressing limitations and further improving the system through clinical validation and optimization of the hardware.

This study provides a robust benchmark for intraoperative brain tumor detection using HSI and machine learning. The high accuracy achieved, along with the detailed methodology and rigorous validation, provide valuable insights for the development of real-time decision support tools in neurosurgery. Future directions include integrating the system into a surgical microscope for enhanced image quality, especially for deep-layer tumors, and conducting extensive clinical trials to evaluate the system's impact on patient outcomes. The findings support the potential of HSI to significantly improve the accuracy and safety of brain tumor resection.

The study's limitations include the challenges associated with acquiring high-quality HSI data during surgery, particularly for deep-layer tumors where optimal focusing can be difficult. Some images with suboptimal acquisition conditions affected the classification results. The dataset, while larger than previously used, is still limited and may not fully represent the diversity of brain tumor types and presentations. Further clinical validation with a larger, more diverse patient population is necessary to fully evaluate the system's generalizability and clinical utility. The preliminary nature of the post-hoc interpretability analysis necessitates further investigation into the model's decision-making processes.