@PHDTHESIS\{IMM1988-01218,
    author       = "K. J. Olsen",
    title        = "Texture analysis of ultrasound images of livers",
    year         = "1988",
    pages        = "162 pp.",
    school       = "{IMSOR,} Technical University of Denmark, {DTU}",
    address      = "Richard Petersens Plads, Building 321, {DK-}2800 Kgs. Lyngby",
    type         = "",
    url          = "http://www2.compute.dtu.dk/pubdb/pubs/1218-full.html",
    abstract     = "It is interesting to compare the results previously obtained and reported in the literature by other research groups and the results in this project. However, one must be very careful when making direct comparisons because many things in the design and analysis may differ.

We have focused on the differentiation between normal liver images and images from livers with diffuse diseases.

For this two group situation the computerized analysis produces hit rates ranging from about 50\% to more than 80\%. Lerski et al. obtained considerably higher hit rates (83-94\%) when differentiating two group situations without mixed cases. However, the statistical treatment of the data must be criticized. The discriminations were neither performed jackknifed nor with calibration and test data set. Furthermore, the reported hit rates seem to be obtained as the maxima of all possible parameter combinations of the eight parameters with highest univariate discriminating capability.

The statistical treatment of the results in Raeth et al. is far better, but one problem is not considered properly. Since they use three images from each liver, the results will be biased, even when they use a jackknifed discrimination. However, the results are good. The overall hit rate, for discriminating between the diffuse diseases fat, hepatitis, cirrhosis/fibrosis and a situation of mixed fat and cirrhosis/fibrosis, is 80\%.

Keeping the differences in statistical methods in mind, the results of the computer analysis in the liver project are quite acceptable. The performance in the visual evaluation lies within the range of the reported experiments. The small superiority of the computer analysis (d-type images) as compared to the visual evaluation, is found in this project as well as in other projects (Raeth et al.)

Since the computer analysis of ultrasound images performs at least as well as the visual evaluation of the images, the combination of the two approaches is likely to improve the precision of the diagnoses. However, the calibration problems with the BK1846 scanner influenced the present dataset so it is not adequate for the design of a final system. In the construction of a final system the following quidelines are advisable from the experiences in this project:

    * Simultaneously with the recording of the patient images an image of a phantom should be recorded. The phantom image must in the computer analysis be used to calibrate for fluctuations in the scanner characteristics.
    * The application of more than one image may improve the precision of the diagnoses. However, the more correlated the images are, the less improvement will be obtained in the diagnoses. In this study a combination of the images from one patient did not yield substantial improvements in the hit rates. Thus it is recommendable to record only one image from each patient. The type d image, a scan of the right liver lobe, gives the best overall hit rate and should be used if only one image is recorded from each patient.
    * The interactive drawing of the contours of the interesting regions is recommendable. In general, it has not been possible to apply the informations from the inhomogeneous parts of the liver in this project. This information is important in visual evaluation of a liver image, but it is apparently difficult to extract the information automaticly. If the inhomogeneous parts are used, the drawing of line between homogeneous and inhomogeneous parts must be done very carefully because the inhomogeneous parts usually are small.
    * There is no significant differences between the two types of contour settings. Thus the standard, which is contour three, may be applied.
    * The histogram parameters and the cartesian {GLCM} parameters seem to contain a very large part of the texture information in the images. A reduction to this set of parameters allows a fast calculation and classification."
}