Colorectal Cancer Detection By Hyperspectral Imaging

Imaging tissues is a vital component in the medical field as it allows an understanding of the anatomical structure and the physiological state of tissues. Current standard of care for gastrointestinal screening typically utilizes endoscopy. Many studies have indicated that early detection is critical in lowering colorectal cancer mortality rates. If colorectal cancer is discovered at an early and localized stage, the 5-year survival rate is 91%; if regional, chances drop to 72%; if distant, chances drop to 14%. Traditional white light endoscopy only provides data in three wavelength bands: red, green, and blue. In contrast, hyperspectral imaging (HSI) is a spectroscopic method that provides data at many wavelength bands. Some HSI configurations provide wavelength information in the UV or IR regions that have historically not been accessible for endoscopy. The additional spectroscopic information provided by HSI allows detection of different tissue types based upon spectral response. The goals of this study were to acquire HSI data from normal and cancerous colorectal tissues and to evaluate the ability of a deep learning system to classify images into categorical groupings. Specimen pairs of colorectal tumors and surrounding noninvolved tissue were obtained through an approved protocol with the Cooperative Human Tissue Network and in accordance with University of South Alabama Institutional Review Board procedures. Hyperspectral image data were obtained where multiple fields of view were imaged from each specimen. Images were then corrected using MATLAB scripts that account for background spectra and correct each spectral image to a uniform flat spectral response. Initial results indicate differences between the spectral characteristics of tissues: cancerous tissues had a lower magnitude of total intensity when compared to noncancerous tissues. Structural differences can be seen when comparing images of cancerous and noncancerous tissues; the noncancerous tissue structure presents a normal crypt structure that would be expected in the colonic mucosa, whereas cancerous tissue structure presents a loss of architecture. Current development focuses on deep learning-based approaches to automatically differentiate between HSI images, noncancerous and cancerous tissue. Future work will focus on expanding the sample dataset to achieve sufficient statistical sampling and translating this technology from a microscope to an endoscope platform for in vivo use. This work was supported by NIH P01HL066299, UL1TR001417, NSF MRI1725937, NSF CNS1953544, Alabama Space Grant Consortium, and Alabama EPSCoR GRF. Drs. Leavesley and Rich disclose financial interest in a start-up company, SpectraCyte LLC, that was formed to commercialize spectral imaging technologies.