Application of decorrelation stretching method to hyperspectral fundus image processing

We have developed a near-infrared hyperspectral imaging system that can acquire both spectral and spatial data covering a 50-degree field at the fundus surface within 5 seconds. Single wavelength band reflectance images with bandwidth of 20 nm have demonstrated that choroidal vascular patterns can be clearly observed as bright images for the central wavelength ranging from 740 to 860 nm, while retinal blood vessels are seen as dark images for that ranging from 740 to 920 nm. It is desirable for clinical use to separate the choroidal vascular patterns image from the retinal blood vessels image. To this end, we have applied the decorrelation stretch to processing of spectral images. We have found the following. Original fundus spectral images have stripes noise. The decorrelation stretch emphasizes the noise and, thus, the noise has to be removed by, for example, DCT (Discrete Cosine Transform) filter beforehand. The choroidal vascular image can be successfully separated from the retinal vascular image. Furthermore, the macular is superimposed on the latter as it should be so from the viewpoint of anatomy. The result suggests that useful information may be extracted by combining hyperspectral images with the decorrelation stretch.