Blood Cells Classification Using Hyperspectral Imaging Technique

Hyperspectral imaging is a relatively new method for identifying blood cells. Except for morphological and texture information in gray images, hyperspectral data contains a lot of spectral signatures which represent chemical analysis of a sample. Therefore, hyperspectrum has an advantage over digital color images due to spectral signatures. With these spectral and spatial features blood cells can be recognized and classified. Over 40 features are extracted from hyperspectral image sequence. These features include spectral pattern traits and similarity measures. To implement blood cell discrimination, a back propagation neural network (BPNN) is proposed in this paper. The connection weights of the BPNN were fixed through the training by a genetic algorithm (GA) which employed two adaptive mechanisms during the evolutional processes. Three-fold cross validation was applied to classify blood cells of given samples. Experimental results demonstrated that the classifier using a BPNN and an adaptive GA was effective. Finally, this paper also described a cursory investigation of the effect of spectral data volume on classification accuracy. Two compressed image series which can be viewed as multispectral series were obtained by systematic sampling from two original hyperspectral series, respectively. Compared to multispectral data, the hyperspectral data with high dimensionality achieved superior accuracy in recognizing blood cells, although requiring greater processing time due to the large amount of data dimension. KeywordsClassification; Blood Cell; Hyperspectral Imaging; BP Neural Network; Genetic Algorithm