Purpose: Stereotactic radiosurgery is a common treatment option for vestibular schwannoma (VS), a slow growing benign brain tumor. The radiation treatment is planned by expert physicians on a brain MRI scan of the patient. Patients who underwent this treatment are monitored by subsequent periodic MRI scans to determine tumor response. The most accurate method to evaluate tumor response is volumetric comparison of the tumor on pre-operative (baseline) and post-operative (follow-up) MRI scans. The accuracy of this evaluation is key to properly assess the efficacy of the treatment and to support clinical decisions for future patient treatment. However, volumetric measurements are currently seldom used since they require manual delineation of the tumor contour by an expert clinician, which is a time-consuming task subject to observer variability. Automatic segmentation of tumors presents challenges which preclude its routine clinical use.
Materials and Methods: We have developed a new algorithm for the automatic segmentation of vestibular schwannoma in follow-up MRI scans using the delineation of the tumor in the baseline MRI scan. The segmentation is performed in three steps: 1) co-registration of the baseline and the follow-up scans; 2) initial segmentation of the tumor in the follow-up scans by k-means clustering using the baseline tumor delineation as a shape prior and; 3) refined segmentation with the 3D Chan-Vese algorithm.
Results: To validate our method, 27 clinical cases of patients with baseline and follow-up MRI scans were collected from the Hadassah University Medical Center. The cases reflect the high variability of the scans resolution and the varying sizes and appearances of the VS tumors. Manual ground truth tumor delineation was obtained for all cases from three experts’ analyses. We quantify the observer ground truth delineation variability with new metrics for the evaluation of accuracy with multiple observers. Our algorithm yields a mean Dice coefficient of 0.88 and 0.89. The Dice coefficients relative to experts’ consensus and mean is 0.93 and 0.94. The tumor growth rate computed by the algorithm yields definite correct clinical decisions in 25 out of 27 cases (one case is borderline).
Conclusion: Our results indicate the high reliability and robustness of our algorithm delineation and its potential clinical use.
