Brain Structure Segmentation
The segmentation of brain structures has always been of interest to the medical community. They provide important information on the shape and volume of structures, used to characterize populations and evaluate the progression of diseases.
Due to manual segmentation being a time consuming task, a number of authomated methods have been proposed in the past years. Traditional atlas based methods can take a whole day to analyze one patient.
Leveraging the recent success of Deep Learning in semantic segmentation, we use Deep Learning models to achieve more robust and fast automated segmentation methods for brain structures.
Diffusion tensor imaging (DTI) is a relatively new modality of Magnetic Resonance Imaging (MRI) able to quantify the anisotropic diffusion of water molecules in highly structured biological tissues. It is unique in its ability to quantify changes in neural tissue microstructure within the human brain non-invasively.
Among the several DTI research fields, we highlight three: segmentation, tractography, and DTI based maps.
DTI-based segmentation aims to delineate regions with similar diffusion characteristics and is a necessary step for performing subsequent quantitative analysis and qualitative visualization. Our group has been working on several DTI-based segmentation and parcellation method. Visit our publication page for more details.
Another important field study of DTI is the proposal of DTI-derived maps, that can be used to highlight or spot specific structures to posterior measurements and analysis. We proposed several maps derived from DTI: the Tensorial Morphological Gradient, suitable for segmentation tasks; the Divergence map, that can be used for landmarking and studying of neuronal fiber organization.
Diffusion Tensor Imaging (DTI) and Morphological Processing
Human Corpus Callosum characterization from Magnetic Ressonance images
The corpus callosum (CC) is the largest white matter structure in the brain composed of axonal fibers crossing both hemispheres. It has a significant role in central nervous system diseases and its volume correlate with severity and extent of neurodegenerative diseases. We are interested in investigating two main topics: CC parcellation and CC signature extraction.
The subdivision of the CC into a given number of regions, known as parcellation, is an important goal with multiple applications, such as studies of CC shape, area, brain connectivity analysis, to studies of the properties of the structure. We also aim to extract the CC signature from magnetic resonance images (MRI) that could be useful for segmentation evaluation, populations comparison, to monitor changes or to identify abnormalities.
Pattern Recognition techniques have been used to develop algorithms to recognize, segment and characterize abnormalities observed in multiple medical image modalities, such as MRI and CT. Our group is currently investigating two applications: human brain white matter lesions using MRI and inflammation/fibrosis in the lung using CT images.
- Analysis of white matter lesion (WML) using MRI: propose an automatic method in order to distinguish between normal and non-normal white matter (recognition task) and also to distinguish different types of lesions based on their etiology: demyelinating or ischemic (classification task). The method combines texture analysis with the use of multiple classifiers.
- Analysis of inflammation/fibrosis in the lung using CT: propose an automatic method that uses computerized tomography (CT) images to automatically segment the lungs (a pre-processing step that is challenging when the pathologies are located in the borders), detect the lesions and distinguish the lung tissue between normal, inflammation and fibrosis.
Abnormalities Analysis using Pattern Recognition Techniques
Mathematical Morphology applied to MRI Segmentation
Mathematical Morphology is an image processing field that analyzes images based on size and shape of structures. Among mathematical morphology tools, our research group works mainly with max-tree and watershed transforms. Max-trees (figure above) represent images through the hierarchical property of thresholds. They can be used to implement connected filters, i.e. filters that do not blur the image, and for segmentation purpouses. The watershed transform (figure below) is a segmentation technique based on flooding of the image. Our group is using max-trees to select robust markers to be used as input to the watesrshed transform. We are currently investigating two applications: brain and carotid segmentation on MR images.