This ratio was used to find the threshold value in all other FLAIR slices and define volumetric ROI of the edema. Although this method may not be very accurate, it is sufficient for our study because histogram features are utilized that consider the pixels in the ROI as an aggregate and thus a few pixel outliers do not affect the resulting features. Using the ROI of the Gd-enhancement and the thickness of each slice, the volume of the Gd-enhanced area was computed. This process was repeated for all of the image series. Then, the relative change in the volume of the Gd-enhancement ALK5 Inhibitor II customer reviews between the baseline and second image series was recorded as a measure of response. This is due to the limitation that only two images series were acquired for some of the patients. Next, a central slice of each volume was chosen for statistical feature extraction.

It should be noted that the tissue characteristics can be reliably measured in the areas without considerable partial volume effects. The central slice has the minimum amount of partial volume and thus can yield the most accurate tissue features. In this step, ROI of the Gdenhanced area was overlaid onto the composite images and their histograms were calculated. Then, a normalization step was applied to them to compensate for the effect of the ROI size. Four histogram features were extracted. Mean and standard deviation, represent average and dispersion of the histogram, respectively. Skewness is a measure of the histogram asymmetry and kurtosis reflects sharpness of the histogram peak. The properties of the last two parameters are illustrated in Figure 3. Altogether, 12 features were extracted from the three composite images. Note that the features are extracted from baseline MR images, whereas the response is measured by comparing the baseline and second series of MR images. We established one-dimensional and multi-dimensional relationships between the proposed features and the extent of response in patients. To this end, single and multiple-regression analyses were done on the results. Prediction equations and the corresponding regression coefficients were derived from these analyses. To control the false discovery rate, we adopted the multiple testing algorithm proposed in.

In addition, leaveone-out cross validation was performed on the data to evaluate the predicted results based on the actual responses of the patients. Also, changes in the volumes of edema and necrosis were evaluated to investigate if they had any relationship with the response of the brain tumor to treatment. Besides these statistical features, we also analyzed the shape and size of necrotic areas of the tumors to see if there were any dependencies between these parameters and the amount of response to the therapy in the patients. This region was selected for this analysis due to its impact on the tumor growth or treatment. In this study, patients with GBM and Gd-enhanced areas were studied to establish a correlation between the response to bevacizumab treatment and features extracted from the structural MR images. Since the Gd-enhanced area of the tumor reflects the most active region of the tumor, the relative change in the volume of this region was considered as a measure of response.