Fewer than one in ten patients diagnosed with the most common type of malignant glioma survives beyond five years, with most living for only 12-18 months after diagnosis^1,2. While complete surgical removal can improve the prognosis, it poses a significant challenge in neurosurgery. The primary difficulty lies in accurately distinguishing tumor cells from healthy brain tissue, particularly at the tumor margins. Incomplete resection often leads to tumor recurrence and poor prognosis for patients. Our recent study, "Towards Machine Learning-based Quantitative Hyperspectral Image Guidance for Brain Tumor Resection", explores the use of hyperspectral imaging (HSI) and machine learning to enhance the precision of malignant glioma surgery and extract crucial tumor characteristics such as molecular markers. These markers could thus be determined intraoperatively and used to guide the surgery.

The Challenge of Glioma Resection

Gliomas, especially high-grade gliomas, are highly invasive, making distinguishing them from normal brain tissue during surgery difficult. Traditional imaging techniques, while helpful, often lack the necessary resolution and specificity to guide surgeons effectively. 5-aminolevulinic acid (5-ALA)-induced protoporphyrin IX (PpIX) fluorescence has improved tumor visibility, but its effectiveness varies, particularly in lower-grade gliomas and low-cellularity tumor portions.

Hyperspectral Imaging: A Game-Changer

Hyperspectral imaging is an advanced technology that provides highly detailed spatially and spectrally resolved information. In other words, unlike conventional imaging, HSI analyzes the entire spectrum of light reflected or emitted by tissue at all points in the image to detect subtle differences in tissue composition. This makes it a powerful tool for distinguishing between tumor and normal tissue, even at the microscopic level, providing detailed spectral information about the examined tissue.

Study Design and Methods

Our study utilized a custom hyperspectral imaging device to capture detailed spectral data from brain tumor tissues in 184 patients. These patients included those with low- and high-grade gliomas, non-glial primary brain tumors, radiation necrosis, and metastases. We collected 891 hyperspectral measurements from these patients and then analyzed them using several machine-learning models, including random forests and multilayer perceptrons.

The hyperspectral imaging device captured fluorescence spectra by illuminating tissue samples with blue light and recording the emitted light using a scientific metal oxide semiconductor (sCMOS) camera through a tunable optical bandpass filter. The data were processed to extract the abundance of five key fluorophores, including the two photo-states of PpIX, NADH, lipofuscin, and flavins.

Overview of hyperspectral imaging and its intraoperative applications. The figure demonstrates various classification possibilities enabled by hyperspectral imaging, including differentiation of tissue types, detection of pathological changes, and real-time surgical guidance.

The figure below outlines the data acquisition and processing steps (1-4) and the classification of the different tumor categories (6, 7). In addition to classification, the endmember abundances can be used for augmented reality overlays (5) and/or statistical analyses (8).

Machine Learning Models and Classification Tasks

The machine learning models were trained to classify tumor type, grade, glioma margins, and IDH mutation status. The key findings from our study are highlighted below:

1. Tumor Type Classification: The random forest classifier achieved an average test accuracy of 87.3% after relabeling certain samples according to the updated WHO classification. This model effectively distinguished between multiple tumor types and healthy tissue.
2. WHO Grade Classification: The multilayer perceptron model performed exceptionally well, with an accuracy of 96.1%, indicating the potential to accurately classify tumor grades intraoperatively.
3. Margin Classification: Identifying the tumor margin and distinguishing between solid tumor, infiltration zone, and reactive brain tissue is crucial for maximizing resection. The margin classification model achieved an accuracy of 85.7%, significantly improving upon previous studies.
4. IDH Mutation Status: The random forest model predicted IDH mutation status with an accuracy of 93%, highlighting the potential for personalized surgical strategies based on genetic characteristics.

Implications for Neurosurgery

The integration of hyperspectral imaging and machine learning represents a significant advancement in the optical imaging of brain tumors. By providing detailed spectral information that surpasses the capabilities of traditional fluorescence imaging, our approach sets the stage for a nuanced distinction between different tumor types.

One significant benefit of this technology is the ability to distinguish between tumor and healthy tissue in real-time accurately. This precision may help surgeons achieve more complete resections, reducing the likelihood of tumor recurrence. Additionally, the ability to classify tumor types and genetic characteristics, such as IDH mutation status, is crucial. For example, during surgery, if a young patient is found to have an IDH-mutated tumor, a surgeon might opt for a less aggressive approach to minimize risk. Conversely, if an IDH wildtype mutation is observed, a riskier approach might be warranted to maximize the patient’s outcome.

By ensuring that more tumor cells are removed during surgery, the risk of recurrence can be significantly reduced, leading to longer patient survival rates. This personalized surgical strategy improves prognosis and tailors postoperative treatment to better meet individual patient needs. Thus, combining hyperspectral imaging, machine learning, and other artificial intelligence methods holds immense promise for revolutionizing the precision and effectiveness of brain tumor surgeries.

Future Directions and Research

While our study's results are promising, several avenues for future research and development remain to be explored. Collecting additional data, especially for less common tumor types, can help improve the accuracy and robustness of the machine-learning models. Additionally, exploring more sophisticated machine learning techniques, such as convolutional neural networks (CNNs), could enhance the ability to analyze spatial information in hyperspectral data cubes.

Integrating hyperspectral imaging into clinical practice will require further validation and refinement. Developing user-friendly interfaces and ensuring seamless integration with existing surgical workflows will be crucial. Moreover, combining hyperspectral imaging with other intraoperative measures, such as ultrasound, neuronavigation, intraoperative neurophysiological monitoring, and even behavioral assessments, could provide a more comprehensive understanding of tumor biology and better guide surgical decisions.

Conclusion

The combination of hyperspectral imaging and machine learning has the potential to revolutionize brain tumor surgery. This approach can enhance the precision and effectiveness of glioma resections by providing detailed, real-time information about tissue composition. Our study demonstrates the feasibility and effectiveness of using hyperspectral imaging to guide surgical decisions, promising improved outcomes for patients with brain tumors.

As we continue to refine and develop these technologies, we move closer to a future where brain tumor surgeries are not only more precise but also tailored to the individual characteristics of each patient's tumor. This represents a significant step forward in the fight against brain cancer, offering new hope to patients and their families.

References

1               Stupp, R. et al. Effects of radiotherapy with concomitant and adjuvant temozolomide versus radiotherapy alone on survival in glioblastoma in a randomised phase III study: 5-year analysis of the EORTC-NCIC trial. Lancet Oncol 10, 459-466 (2009). https://doi.org:10.1016/S1470-2045(09)70025-7

2               Stupp, R. et al. Radiotherapy plus concomitant and adjuvant temozolomide for glioblastoma. N Engl J Med 352, 987-996 (2005). https://doi.org:10.1056/NEJMoa043330