3144 - Single-Pixel Approach Achieving Ultrasensitive In Vivo Bioluminescence Imaging for Cancer Research

Purpose/Objective(s): The ability to detect and monitor individual cells in vivo is critical for understanding cancer progression and assessing treatment response in preclinical settings. Bioluminescence imaging (BLI) is commonly used for this purpose; however, conventional optical imaging (CI), which relies on 2D array detectors, faces significant challenges in detecting biological events involving small cell populations or weak in vivo signals. To address this, we propose a highly innovative approach that combines a single-pixel framework with advanced machine learning algorithms to achieve ultrasensitive in vivo BLI. We conduct a comprehensive investigation on single-pixel imaging (SPI) for low BL detection including theoretical studies, imaging sensitivity and quality, and ultra-low signal detection in both in vivo tumor models and phantom, comparing its performance to that of CI.

Materials/Methods: The SP approach uses a light modulator, digital micromirror device (DMD), to generate spatial patterns that modulate the optical signal emitted by the subject. An SP detector captures the modulated signals corresponding to each pattern, which are then computationally reconstructed into a 2D image. The key advantages of SPI lie in its unique light collection mechanism and the use of advanced reconstruction algorithms, such as total variance and FFDNet, enabling significantly higher imaging sensitivity compared to CI. We theoretically determined the required number of modulated patterns for SPI to reconstruct a K-sparse signal and derived conditions under which SPI outperforms CI in terms of signal-to-noise ratio (SNR). To further evaluate the imaging sensitivity of SPI and CI as a function of SNR, we simulated an SP detector using a CCD camera with pixel binning along with the DMD and compared its performance to that of a CCD camera used for CI. Additionally, we examined SPI's performance in detecting BL-H460 lung cancer cells (1–100 cells) compared to CI in phantom. For in vivo studies, we intracranially injected H460 cells to mimic brain metastases at different cell numbers and assessed SPI's ability to detect early malignancies.

Results: Using artificial BL light sources, we show not only 16-fold higher SNR of SPI compared to CI, but also SPI can reconstruct images as low as 5.7×10^-6 lux with FFDNet, while CI shows noisy images, consistent with our theoretical analysis. Moreover, SPI shows significant improvement in image contrast, as high as 5-fold, especially in low-signal or high noise conditions i.e. SNR = 1/3, compared to CI. We also successfully detected a single H460 cell implanted at a depth of 7 mm in a tissue-mimicking mouse phantom, while CI failed to detect the signal even after 1000 sec. of acquisition time.

Conclusion: The SP approach will redefine in vivo imaging by enabling unparalleled imaging sensitivity, overcoming fundamental limitations in current cancer research. This technology is expected to have an immediate impact on understanding cancer progression and therapeutic response.