Segmentation in dermatological hyperspectral images: dedicated methods

Abstract

BackgroundSegmentation of hyperspectral medical images is one of many image segmentation methods which require profiling. This profiling involves either the adjustment of existing, known image segmentation methods or a proposal of new dedicated methods of hyperspectral image segmentation. Taking into consideration the size of analysed data, the time of analysis is of major importance. Therefore, the authors proposed three new dedicated methods of hyperspectral image segmentation with special reference to the time of analysis.MethodsThe segmentation methods presented in this paper were tested and profiled to the images acquired from different hyperspectral cameras including SOC710 Hyperspectral Imaging System, Specim sCMOS-50-V10E. Correct functioning of the method was tested for over 10,000 2D images constituting the sequence of over 700 registrations of the areas of the left and right hand and the forearm.ResultsAs a result, three new methods of hyperspectral image segmentation have been proposed: fast analysis of emissivity curves (SKE), 3D segmentation (S3D) and hierarchical segmentation (SH). They have the following features: are fully automatic; allow for implementation of fast segmentation methods; are profiled to hyperspectral image segmentation; use emissivity curves in the model form, can be applied in any type of objects not necessarily biological ones, are faster (SKE—2.3 ms, S3D—1949 ms, SH—844 ms for the computer with Intel® Core i7 4960X CPU 3.6 GHz) and more accurate (SKE—accuracy 79 %, S3D—90 %, SH—92 %) in comparison with typical methods known from the literature.ConclusionsProfiling and/or proposing new methods of hyperspectral image segmentation is an indispensable element of developing software. This ensures speed, repeatability and low sensitivity of the algorithm to changing parameters.

Hyperspectral imaging in medicine: image pre‐processing problems and solutions in Matlab

“…due to the increase in the measurement time of a given sample. By providing a perpendicular position of the camera relative to the object and even illumination, calibration can be simplified to direct registration of the object in the area, where, for example, two patterns, white and black, will be placed in the extreme corners [2]. These areas may be of any size but not less than half the size of the median filter mask (if used).…”

“…There are also other methods of hyperspectral image calibration [2], but their application in practice is much less cumbersome. One of them is the use of black and white patterns located for several columns on the image edge.…”

“…Sensitivity range of the camera and its bandwidth are individually selected, most often at the time of buying the camera [1]. Depending on the individual choice of these parameters and the selected scan area, the image L GRAY (m, n, i) is registered, where m-row, n-column, i-number of a 2D image in a sequence [2]. The image L GRAY (m, n, i) has an M × N × I resolution, where M-number of rows, N-number of columns, I-number of 2D images, equal to the number of different wavelengths λ during acquisition.…”

“…The image L GRAY (m, n, i) has an M × N × I resolution, where M-number of rows, N-number of columns, I-number of 2D images, equal to the number of different wavelengths λ during acquisition. In relation to the known types of software (FreeMat, Matlab, Scilab, Octave) that use raster graphics and nomenclature used in hyperspectral imaging, there occur synonyms that are used later in this paper, lines (rowssymbol m), samples (columnssymbol n) and bands (image number for a given wavelengthsymbol i) [3]. The method for saving this type of image on a computer is different and is closely dependent on the type of camera used.…”

“…Specim, these are *.dat or *. raw files [3]. Both of these file formats are closely associated with the specificity of the camera operation itself.…”