Metasurface-enhanced terahertz imaging for glioblastoma in orthotopic xenograft mouse model combined with neural network decision making.

Terahertz (THz) optical sensing and imaging offer significant potential in a range of biological and medical applications owing to their low-energy, non-ionizing nature, and ultra-broadband spectral information, which includes numerous molecular fingerprints. However, conventional THz imaging suffers from limited contrast and low absorption cross-section in biological tissues. Recent advances in terahertz sensing platforms, facilitated by various metasurfaces, have addressed these limitations by enhancing the sensitivity and selectivity of optical detection and imaging. This study presents an advanced label-free terahertz imaging technique that leverages a metasurface to enhance image contrast. We applied this method to image glioblastoma model mouse brain tissues. To identify cancerous regions clearly, the complex refractive indices across the brain tissues were determined using a finite element method simulation. Furthermore, the strong resonance features of the metasurface facilitate correlation-based learning in neural networks. We employed a convolutional neural network to segment cancer boundaries using the metasurface-enhanced imaging data. Glioblastoma regions were identified with an accuracy of over 99 %, by using fluorescence-labeled images as the training data for the neural networks. This study highlights the critical role of metasurfaces in fundamentally enhancing terahertz wave-matter interactions and how integration with neural networks enables highly sensitive cancer detection. This paves the way for the clinical applications of terahertz imaging technologies in medical diagnostics.