NI-76 * COMPUTER EXTRACTED ORIENTED TEXTURE FEATURES ON T1-GADOLINIUM MRI FOR DISTINGUISHING RADIATION NECROSIS FROM RECURRENT BRAIN TUMORS

Abstract

OBJECTIVE: Differentiating radiation necrosis (RN), a delayed response to radiation, from recurrent brain tumors is a clinically challenging problem both for primary and metastatic patients; due to their similar appearances on gadolinium-contrast T1-weighted MRI (Gd-T1). Biopsy/ resection is currently the only method for definitive diagnosis. To date no radiographic features have been definitely identified that can distinguish between RN and recurrent brain-tumors on T1-w MRI. We present a new gradient orientation texture feature that aims to capture the differential spatial patterns between RN and recurrent brain-tumors, the hypothesis being that recurrent tumors have more heterogeneity in appearance and hence the texture gradients are more randomly oriented compared to RN. METHOD AND MATERIALS: We analyzed 42 immediate presurgical Gd-T1 MRI studies, 20 primary (10 RN, 10 recurrent tumors) and 22 metastatic (10 RN, 12 recurrent brain-tumors). All patients underwent resection or multiple biopsies of the recurrent imaging lesion. An expert neuropathologist, blinded to the original diagnosis and type of radiation, established the diagnosis of RN versus recurrence. Orientation of every pixel within the lesion is obtained by calculating intensity gradient of every pixel based on the proximal pixels in X and Y-direction. We then compute entropy of these orientations for every pixel on an image. The entropy values across RN and recurrent tumors are curated and evaluated for both pathologies. RESULTS: The mean normalized entropy values were found to be: 0.95 + /-0.13 (recurrence metastasis), 0.91 +/- 0.08 (RN metastasis), and 0.89 + /- 0.09 (recurrence primary), and 0.87 +/- 0.1 (RN primary) within the lesion area, with the accuracy of 80% for distinguishing RN and recurrence in metastatic and 77% for primary brain tumor patients. CONCLUSION: radiation necrosis and recurrence manifest significantly different textural oriented gradients for both primary and metastatic brain tumors, features that are not visually discernible but resolvable by a computer.