Segmentation of diffuse reflectance hyperspectral datasets with noise for detection of Melanoma

We present a segmentation algorithm that allows optical properties to be extracted from diffuse reflectance hyperspectral datasets with a speedup of three orders of magnitude when compared to current methods. Such data could be used for the detection of melanoma. The algorithm first performs dimensionality reduction using principal component analysis, and then the image is segmented using k-means clustering. The mean spectrum from each cluster is then calculated and can be used to extract chemical information. By reducing the number of spectra to be analyzed, extraction of physiological information can be achieved three orders of magnitude faster than methods requiring the analysis of every spectrum in the hyperspectral dataset. The effect of noise on the ability of the algorithm to accurately segment images was tested using digital phantoms, for which the noise level was under the control of the investigators. The analysis showed a linear relationship between the level of noise and the smallest difference in scattering that the algorithm was able to accurately detect and segment. This finding can be used to determine the maximum amount of noise in the imaging system that will still allow detection of the difference in optical properties between non-melanoma and melanoma.