List of communications


Glioblastoma tissue-guided segmentation through unsupervised structured classification

Javier Juan-Albarracín, Grupo de Informática Biomédica (IBIME). Instituto de Aplicaciones de las Tecnologías de la Información y de las Comunicaciones Avanzadas (ITACA). Universitat Politècnica de València. Valencia. Spain; Elies Fuster-Garcia, Grupo de Informática Biomédica (IBIME). Instituto de Aplicaciones de las Tecnologías de la Información y de las Comunicaciones Avanzadas (ITACA). Universitat Politècnica de València. Valencia. Spain; Juan M. García-Gómez, Grupo de Informática Biomédica (IBIME). Instituto de Aplicaciones de las Tecnologías de la Información y de las Comunicaciones Avanzadas (ITACA). Universitat Politècnica de València. Valencia. Spain


The early identification of the tissues involved in the Glioblastoma is crucial to make decisions that can improve the patient survival. Nowadays, the identification of these tissues is performed manually although it involves a tedious, time-consuming and biased process. Hence, several approaches have been proposed in the literature to address the problem automatically. Most of these approaches usually arise from the supervised learning standpoint, however such paradigm suffers from several limitations such as: 1) the huge cost of retrieving a manually annotated corpus from where to learn the models of the tissues, 2) the sensibility of the models to changes in the MRI protocols that can distort the data, or 3) the over-fitting problem during the learning process.

In contrast, a recently published unsupervised approach evaluated with the BRATS 2013 dataset was able to avoid these limitations while obtaining comparable results to supervised approaches. Our work presents a significant improvement of this method. Instead of a global classification of the brain, we propose a tissue-guided approach that segments each tissue independently using only its most discriminant MR images and features. We also evaluated our method with BRATS2013 data considering both structured and non-structured classification algorithms. The proposed approach allows us to remove the post-processing stage, improving the computational performance of the method and making it more transparent and stable.

image
Format: Poster

Organized by