Validation of subpixel target detection and linear spectral unmixing techniques on short-wave infrared hyperspectral images of collagen phantoms.

Significance: We used three-dimensionally printed experimental molds and designed lard (lipid)-collagen mixed phantoms to simulate biological tissues to verify the practicality and accuracy of short-wave infrared (SWIR) hyperspectral imaging (HSI; 900 to 1700 nm), subpixel target detection (STD), and linear spectral unmixing (LSU). We provide a foundation for future development, validation, and reproducibility of hyperspectral image-processing techniques.

Aim: We aim to verify the use of SWIR HSI in bionic tissue phantoms. Second, we focus on the accuracy of STD and spectral unmixing techniques in hyperspectral image processing. Finally, the penetration ability of the technology and its applications at various depths and concentrations are explored.

Approach: All experiments were conducted using an SWIR (900 to 1700 nm) HSI sensor. Collagen phantoms of different thicknesses were created to test the penetration abilities. Lard (lipid) was embedded at different depths in the phantoms for STD, whereas LSU was performed on phantoms with varying collagen concentrations. The methods used included constrained energy minimization to detect the lard target and fully constrained least squares (FCLS) to estimate the abundance of collagen phantoms.

Results: SWIR HSI effectively penetrated the collagen phantoms. Specifically, STD techniques can accurately detect the presence of lard (lipids) at depths of 7 to 20 mm in the collagen phantoms. Even at a depth of 68 mm, the detection accuracy was 0.907. Moreover, in the LSU analysis, the FCLS method accurately estimated the abundance of collagen phantoms at different concentrations, with a correlation coefficient of 0.9917, indicating high accuracy across different concentrations.

Conclusions: This study demonstrated that SWIR HSI is highly accurate for deep target detection and LSU. This technology has great potential for use in future noninvasive biomedical diagnostic models. Collagen phantoms are valuable tools for validating HSI algorithms and provide a solid foundation for clinical applications.