Cameron May¹ (), Karthik Kasi, Leo Kobayashi, Bevil Conway, Joseph Pare; ¹National Institutes of Health

Hyperspectral imaging (HSI) is emerging as a powerful tool in medical diagnostics for its ability to capture and analyze data from across the electromagnetic spectrum. In contrast with traditional RGB imaging which is limited to just three-color channels, HSI allows for more color channels across a broader spectrum. HSI allows for the identification of unique spectral signatures based on how human tissues interact with different wavelengths of light. In medicine, HSI has been applied to distinguish between pathological and non-pathological states such as kidney stones or cancers. However, most studies focus on pixel-level classification (e.g., detecting cancer in specific pixels), rather than whole-image classification methods commonly used in convolutional machine learning (e.g., determining if an entire image shows signs of disease). We address this question by determining the extent to which whole image classification can discriminate between edematous or non-edematous legs as a proof of principle. Edematous legs are associated with different levels of perfusion which should manifest in different spectra observed with HSI. Here, we test this prediction. We used a push-broom hyperspectral camera (SOC-710, Surface Optics Corporation) to capture images and assess the classification potential of this method. 68 Images were captured from 27 emergency room patients over 4 months. We generated 5x5 pixel patches labeled as edematous or non-edematous. Using principal components analysis for unsupervised dimensionality reduction, we achieved a 66% classification accuracy with support vector machines, utilizing a linear kernel. This analysis demonstrates HSI's potential to expand access to medical imaging by offering a more affordable and portable alternative to conventional methods, particularly benefiting under-resourced clinics.