Classification of JPEG2000 compressed CT images

The paper aims on an experimental analysis of the impact of JPEG2000 compression to the classification of DICOM liver tissue images. The classification algorithm computes a depth-three threshold circuit from a sample set S of 400 positive (abnormal findings) and 400 negative (normal liver tissue) examples by a local search strategy. The local search is based on simulated annealing where the neighborhood relation is determined by the classical perceptron algorithm. The examples are fragments of DICOM CT images of size n=14161=119×119 pixels. The images are encoded with different parameter settings of JPEG2000 and a given buffer size, i.e., the size of the resulting encoded image is used as parameter. The algorithm is trained with decoded images and tested on sets of 100+100 examples (disjoint from the learning set) of decoded and original images. Results show that the classification is robust against different levels of compression and performs a correct classification of about 97%.