1 m f I , f I ) ( + − ∈ x I where I(x) is the intensity value of the pixel at position x. The first iteration is completed after all the valid pixels are included in the region. In the next iteration, mean and standard deviation of the intensity values are recomputed using all the pixels currently included in the region. A new intensity range is calculated from these values and the neighbors of the current region are tested for their inclusion in the region and are added to the region as appropriate. The iterations are continued until no more pixels are added or the maximum number of iterations is reached. The segmentation of hyperintense lesions in FSE-T2 images is known to be a difficult task because the lesions and CSF can have the same intensity values in these images (Fig. 4(a)). We addressed this problem by using the mask obtained by first segmenting CSF from the T1 and then applying it as an exclusion mask. Results: The algorithm was used to successfully segment hyperintense lesions in FLAIR and T2 images and both hyper-and hypointense lesions separately in T1 contrast enhanced images. Fig 1 shows a 3D view of the segmented lesions on a FSE-T2 image. Fig. 2 shows the volume comparison of the hyperintense lesions on FLAIR images by automated and manual tracing and Fig.3 shows a similar comparison of the results for the contrast enhanced lesion in T1 images for two patients after surgical removal of active tumor. For validating segmentation of FSE images the similarity coefficient (2), which is a measure of overlap between automatically segmented volume and manual segmented data was calculated and was observed to be 81.7 ± 5.88 (n=7). The overlap matrices (2) between the manually and automatically segmented volumes in 30 GlioblastomaMultiformes was found to be 80.09 ± 8.5 for T2 lesions and 74.66 ± 9.31 for contrast enhanced tumors. Time required for the execution of the algorithm was a few seconds per case in a 2.6 GHz Intel Xeon processor. Discussion: Our findings show that both meningiomas and gliomas can be accurately segmented by means of automated processing, as shown in Fig. 4. Segmentation of lesions with complicated shapes that are difficult to analyze using manual segmentation is possible with this algorithm. In cases where the tumor was very close to the skull or eye region we did a skull stripping prior to running the segmentation which excluded the extraneous objects even before lesion segmentation started. Although we tried refining the segmentation by applying Geodesic level sets (3), the segmentation was not improved considerably. A similar refinement using a fuzzy analysis method (4) was found to increase the computations without considerably improving the segmentation results. For this class of images, we therefore decided to avoid computationally burdensome refinement operations. Conclusions: The results show that volume measurements obtained using this method are in good agreement with manually segmented data. The implementation of the method is being incorporated as part of the routine evaluation of follow-up data for patients with brain tumors in our institution.