Brain tissue classification in hyperspectral images using multistage diffusion features and transformer.

Area of Science:

• Medical imaging
• Optical engineering
• Artificial intelligence

Background:

• Accurate brain tumor boundary identification is crucial for effective surgical resection and minimizing complications.
• Hyperspectral imaging (HSI) offers detailed spectral information for enhanced intraoperative tissue classification.
• Current HSI analysis methods face challenges with long processing times, high computational costs, and the need for model retraining.

Purpose of the Study:

• To develop a novel, efficient framework for hyperspectral image (HSI) analysis in surgical applications.
• To overcome the limitations of existing HSI classification methods, including retraining requirements and computational expense.
• To improve intraoperative tissue classification accuracy using spectral-spatial features.

Main Methods:

• Proposed MedDiffHSI framework combining diffusion models and transformer techniques for unsupervised spectral-spatial feature extraction.
• Utilized a pretrained denoising U-Net to extract intermediate multistage features from diffusion model outputs.
• Implemented a spectral-spatial attention module to enhance feature representation and a transformer-based classifier with weighted majority voting (WMV) for classification.

Main Results:

• MedDiffHSI demonstrated state-of-the-art performance in HSI classification using minimal training samples (5%).
• The framework achieved superior results compared to existing approaches on in vivo brain tumor datasets.
• Validation extended to breast cancer HSI datasets, showing robust performance across different tissue types.

Conclusions:

• MedDiffHSI offers an efficient and effective solution for intraoperative tissue classification using hyperspectral imaging.
• The unsupervised learning approach reduces computational complexity and eliminates the need for model retraining.
• This method holds significant potential for improving surgical decision-making and patient outcomes in neurosurgery and beyond.