Automatic extraction of tumors from multiple MR images with self-organizing maps

In MR images, the contrast between tumors and surrounding soft tissues is not always clear, and it may be difficult to determine the tumor region. In this report, we propose a method for the automatic and objective extraction of tumors based on the correlations among multiple MR images. First, a map reflecting the correlations of three types of MR images (Gd-enhanced, T1-weighted, and T2-weighted images) is created by training of Self-Organizing Maps (SOM). Second, the SOM are grouped into a number of clusters determined in advance, and the original MR images are divided into clusters according to the clustered SOM. Finally, the tumor region in the clustered MR images is refined by reclassification, improving the accuracy of extraction. This method was applied to 10 cases in a clinical study, and in 8 of these cases, the tumor could be distinguished from other regions as an independent cluster. The proposed method is expected to be useful for the automatic extraction of tumors in MR images. (author)