A new approach for texture characterization of tumors based on a 2D S-Transform. Previous methods have already been developed using transforms such as the wavelet transform or the Gabor transform. However, these transforms are either not frequency invariant (wavelet transform) or have a fixed resolution (Gabor transform) To solve these problems we employ a 2D S-Transform. The S-Transform, a generalization of the Short-time Fourier transform, provides a time-frequency distribution of a signal. Therefore one can obtain the frequency content of a pixel or of a tumor ROI by averaging pixel spectrums over the tumor . A tool has been developed that computes the S-Transform in real time for a pixel and in 2 or 3 seconds for a tumor, while previous methods take much longer time. To quantify the tumor texture we compute statistics based on pixel spectrums. The first statistic, a texture curve, is the frequency vs its average power at each pixel or over the entire tumor. The second statistic, for a tumor, is a KL-divergence to calculate the deviation of histograms, obtained at each frequency, from a normal distribution. Finally, a map of the area under the texture curve for a band of frequencies shows the average power at each pixel of the tumor. First experiments on images of 20 tumor-bearing patients (10 for the training set, 10 for the test set) using the texture curves allowed us to classify homogeneous and heterogeneous tumors with an accuracy of around 80%.
