The development of brain tumor segmentation techniques based on multi-spectral MR image data has relevant impact on the clinical practice via better diagnosis, radiotherapy planning and follow-up studies. This task is also very challenging due to the great variety of tumor appearances, the presence of several noise effects, and the differences in scanner sensitivity. This paper proposes an automatic procedure trained to distinguish gliomas from normal brain tissues in multi-spectral MRI data. The procedure is based on a random forest (RF) classifier, which uses 80 computed features beside the four observed ones, including morphological ones, gradients, and Gabor wavelet features. The intermediary segmentation outcome provided by the RF is fed to a twofold post-processing, which regularizes the shape of detected tumors and enhances the segmentation accuracy. The performance of the procedure was evaluated using the 274 records of the BraTS 2015 train data set. The achieved overall Dice scores between 85-86% represent highly accurate segmentation.

The main drawback of magnetic resonance imaging (MRI) represents the lack of a standard intensity scale. All observed numerical values are relative and can only be interpreted together with their context. Before feeding MRI data volumes to supervised learning segmentation procedures, their histograms need to be registered to each other, or in other words, they need a so-called normalization. The most popular histogram normalization technique used to assist brain MRI segmentation is the algorithm proposed by Ny úl et al in 2000, which aligns the histograms of a batch of MRI volumes without paying attention to possible focal lesions that might distort the histogram. Alternately, some recent works applied histogram normalization based on a simple linear transform, and reported achieving slightly better accuracy with them. This paper proposes to investigate, which is the most suitable method and parameter settings for histogram normalization to be performed before the segmentation of brain MRI images, separately in the cases of absence and presence of focal lesions.

Ensemble learning methods are frequently employed in medical decision support. In image segmentation problems the ensemble based decisions require a postprocessing, because the ensemble cannot adequately handle the strong correlation of neighbor voxels. This paper proposes a brain tumor segmentation procedure based on an ensemble cascade. The first ensemble consisting of binary decision trees is trained to separate focal lesions from normal tissues based on four observed and 100 computed features. Starting from the intermediary labels provided by the first ensemble, six local features are computed for each voxel that serve as input for the second ensemble. The second ensemble is a classical random forest that enforces the correlation between neighbor pixels, regularizes the shape of the lesions. The segmentation accuracy is characterized by 85.5% overall Dice Score, 0.5% above previous solutions.

No abstract