Influence of MRI acquisition protocols and image intensity normalization methods on texture classification

Texture analysis methods quantify the spatial variations in gray level values within an image and thus can provide useful information on the structures observed. However, they are sensitive to acquisition conditions due to the use of different protocols and to intra- and interscanner variations in the case of MRI. The influence was studied of two protocols and four different conditions of normalization of gray levels on the discrimination power of texture analysis methods applied to soft cheeses. Thirty-two samples of soft cheese were chosen at two different ripening periods (16 young and 16 old samples) in order to obtain two different microscopic structures of the protein gel. Proton density and T(2)-weighted MR images were acquired using a spin echo sequence on a 0.2 T scanner. Gray levels were normalized according to four methods: original gray levels, same maximum for all images, same mean for all images, and dynamics limited to micro +/- 3sigma. Regions of interest were automatically defined, and texture descriptors were then computed for the co-occurrence matrix, run length matrix, gradient matrix, autoregressive model, and wavelet transform. The features with the lowest probability of error and average correlation coefficient were selected and used for classification with 1-nearest neighbor (1-NN) classifier. The best results were obtained when using the limitation of dynamics to micro +/- 3sigma, which enhanced the differences between the two classes. The results demonstrated the influence of the normalization method and of the acquisition protocol on the effectiveness of the classification and also on the parameters selected for classification. These results indicate the need to evaluate sensitivity to MR acquisition protocols and to gray level normalization methods when texture analysis is required.