Imaging Through Scattering Tissue Using Near Infra-Red and a Convolutional Autoencoder.

Area of Science:

• Biomedical Optics
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate tumor margin delineation is crucial for effective cancer surgery.
• Current imaging techniques (MRI, CT, fluorescence) have limitations like cost, accessibility, and intraoperative constraints.
• There is a need for advanced, non-invasive imaging solutions for subsurface visualization.

Purpose of the Study:

• To develop and evaluate a low-cost, non-invasive subsurface imaging approach using near-infrared (NIR) illumination and deep learning.
• To reconstruct hidden structures from highly scattered surface images.
• To assess the potential for clinical applications, such as tumor margin detection.

Main Methods:

• A controlled experimental setup using a tissue-mimicking chicken phantom.
• Structured patterns were concealed beneath the phantom and imaged with a NIR-sensitive camera under halogen illumination.
• A convolutional autoencoder (U-Net architecture) was trained on ~10,000 paired samples for image reconstruction.

Main Results:

• The deep learning model achieved strong reconstruction performance, outperforming conventional Wiener filtering.
• Key performance metrics included a peak signal-to-noise ratio (PSNR) of 20.14 dB, structural similarity index (SSIM) of 0.92, and feature similarity index (FSIM) of 0.94.
• Accurate recovery of subsurface shapes was demonstrated, with minor smoothing artifacts.

Conclusions:

• The combination of NIR imaging and deep learning offers a viable method for safe, rapid, and cost-effective subsurface visualization.
• This approach shows significant potential for advancing tumor margin detection in clinical settings.
• Further development is needed to improve generalization to diverse samples.