Deep learning-based hyperspectral image correction and unmixing for brain tumor surgery

Area of Science:

• Medical imaging
• Neurosurgery
• Artificial intelligence

Background:

• Hyperspectral imaging (HSI) aids fluorescence-guided brain tumor resection by visualizing tissue differences.
• Current HSI methods struggle with accuracy due to uncorrected optical and geometric tissue variations.
• Improved accuracy in HSI can lead to better patient outcomes in neurosurgery.

Purpose of the Study:

• To develop and evaluate deep learning (DL) models for correcting and unmixing HSI data in fluorescence-guided neurosurgery.
• To improve the accuracy of protoporphyrin IX (PpIX) concentration estimation.
• To address limitations of classical methods in handling heterogeneous tissue properties.

Main Methods:

• Proposed two DL models: one supervised (trained with PpIX concentration labels) and one semi-supervised.
• Evaluated models on phantom and pig brain data with known PpIX concentrations.
• Assessed generalization to human data and compared performance against classical approaches.

Main Results:

• Supervised DL model achieved Pearson correlation coefficients of 0.997 (phantom) and 0.990 (pig brain).
• Semi-supervised DL model achieved Pearson correlation coefficients of 0.98 (phantom) and 0.91 (pig brain).
• Classical methods yielded lower correlations (0.93, 0.82).
• Semi-supervised model demonstrated superior generalization to human data, with a 36% lower false-positive rate for PpIX detection.

Conclusions:

• Deep learning models significantly improve the accuracy of hyperspectral fluorescence-guided neurosurgery.
• The semi-supervised DL approach shows promise for real-world clinical application due to better generalization.
• These advancements can enhance visualization and precision in brain tumor resection, potentially improving patient outcomes.