Tumor segmentation from single contrast MR images of human brain

Detection of brain tumors may help in diagnosis, therapy planning and treatment planning. Brain tumor detection in scout scan may also help in identify the surrounding tissue of the pathology and thus help in imaging parameter optimization. We propose an automatic method to segment brain tumor on one single T2W image. The proposed method for the brain tumor segmentation consists of three steps. In the first step, we normalize the image intensity and register to a standard brain space. Second, we perform a pixel-wise classification using a random forest classification method. For each pixel, 198 features are extracted, including multiscale intensity-based features, multiscale template-based features, multiscale shape based methods, multiscale texture-based features as well as context-aware features. After the pixel-wise classification, we further exclude false positives in a morphological way. Each independent connected objects from the second step, which either has the maximum volume or has a volume over 10 cm³, is considered as true tumors. The classifier is trained and evaluated using a leave-one-out framework on a publically available database. Our method successfully detects out tumors in 28 out of 30 dataset (successful rate = 93.3%). The accuracy for the pixel-wise classification is averagely 97.5% for low grade tumors and 96.7% for high grade tumors. The final segmentation is evaluated using Dice Similarity Coefficient, which is 78.8% for low grade tumors and 83.0% for high grade tumors.