Automatic segmentation of subdural hematomas using a computational technique based on smart operators

This paper proposes a non-linear computational technique for the segmentation of subdural hematomas (SDH), present in 4 multilayer computed tomography brain imaging databases. This technique consists of 3 stages developed in the three-dimensional domain. They are: pre-processing, segmentation and quantification of the volume occupied by each of the segmented SDHs. The pre-processing stage is divided into two phases. The first one, called the definition of a volume of interest (VOI), a band thresholding algorithm is used which allows, fundamentally, to isolate the SDH considered from the rest of the surrounding anatomical structures. In the second phase, filtering, a bank of computational algorithms is applied to reduce the impact of the artifacts and attenuate the noise present in the images. The algorithms that make up this phase are: the morphological erosion filter (MEF) and the median filter (MF). On the other hand, during the segmentation stage a clustering algorithm, called Region Growing (RG), is implemented and it is applied to the pre-processed images. The RG requires for its initialization a seed voxel whose coordinates are obtained automatically through the training and validation of intelligent operators based on support vector machines (SVM). Due to the high sensitivity of the RG to the location of the seed, the SVMs are implemented as highly selective binary classifiers. On the other hand, in order to compensate for the effect of the MEF the SDH, which has been preliminarily segmented, is submitted to the application of a morphological dilation filter of binary type (MDF). To make value judgments about the performance of the proposed technique, the SDH dilated segmentations, obtained automatically, and the SDH segmentations, generated manually by a neurosurgeon, are compared using the Dice coefficient (Dc). The combination of parameters linked to the highest Dc value, allows to establish the optimal parameters of each of the computational algorithms that make up the proposed nonlinear technique. The obtained results allow to report a Dc superior to 0.87 which indicates a good correlation between the manual segmentations and those produced by the computational technique developed. Finally, as an immediate clinical application, considering the automatic segmentations, the volume of each hematoma is calculated considering both the voxel size of each database and the number of voxels that make up the segmented hematomas.