| Summary: | Magnetic Resonance Imaging is the preferred imaging modality for assessing brain tumors, and segmentation is necessary for diagnosis and treatment planning. Thus, robust automatic segmentation methods are required. Machine learning proposals where the model is learned from data are quite successful. Hierarchical segmentation approaches firstly segment the whole tumor, followed by intra-tumor tissue identification. However, results comparing it with single stages approaches are needed, as state of the art results are also achieved by all-at-once strategies. Currently, fully convolutional networks approaches for segmentation are very efficient. In this paper, a hierarchical approach for brain tumor segmentation using a fully convolutional network is studied. The evaluation is performed on the Brain Tumor Segmentation Challenge 2013 dataset, and we report the metrics Dice Score Coefficient, Positive Predictive Value, and Sensitivity. Results show benefits from segmenting the complete tumor first, over all tissues in one stage. Moreover, the tumor core also benefits from such approach. This behavior may be justified by the high data imbalance observed between tumor and normal tissues, which is mitigated by considering the tumor as a whole.
|
|---|