Classification of soft cheese image texture for different MR protocols
Résumé
Texture analysis methods are very useful tool used to describe the grey level distribution in an image. These methods are widely used for analysis of MRI images in medicine and food industry, especially for classification of different textures. The objective of this paper is to present analysis results concerning the influence of the different MR protocols and image normalisation on the classification of cheese image texture. Soft cheese samples were chosen at two different ripening periods in order to have two different microscopic structures of the protein gel. The analysed data set contained 32 samples, 16 of each class. MR images were acquired using a 0.2T Siemens scanner. Two pulse sequences were used: a spin-echo sequence to produce proton-density, T2-weighted and T1-weighted images, a gradient echo (FLASH) sequence to produce magnetic susceptibility-weighted images. The analysis was performed for original grey levels and for normalised image (m±3s ). Texture descriptors were computed for: co-occurrence matrix (220 parameters), run length matrix (20), gradient matrix (5) auto-regressive model (5), wavelet transform (32), morphological decomposition (40). The features with the highest Fisher coefficient were selected and used for classification performed by two classifiers: 1- nearest neighbour (1-NN) and nonlinear discriminant analysis (NDA). Applied texture analysis was able to discriminate between two different gel structures, but classification results depend on the MR protocol and image normalisation. In this study normalisation provides better classification, which may remove between scanners variations, however more study on this subject is required.