Michal Sofka presents Detecting and Segmenting Organs in 3D Medical Images
On 2012-06-05 11:00
at G205, Karlovo náměstí 13, Praha 2
Detecting jointly multiple organs and anatomical structures in 3D medical
images
is difficult to obtain in most practical situations due to the large parameter
space and due to data variations. We address this problem by a Hierarchical
Detection Network (HDN) that detects the structures sequentially, one-by-one.
The interdependence of structures poses and prior information embedded in our
domain of medical images results in better performance than detecting the
structures individually. The posterior distribution of each structure pose is
approximated at every step by sequential Monte Carlo sampling. The samples are
propagated within the sequence across multiple objects and hierarchical levels.
The most reliable samples are used to initialize a multi-organ segmentation
system. An example of this procedure will be shown on segmenting pathological
lung in CT images. The lung pose initializes the detection of a set of
automatically-selected landmarks which typically lie near geometric structures
both within and outside the lung (e.g. ribs and spine). These stable landmarks
are used to robustly initialize a coarse statistical model of the lung shape
and
help to cope with higher attenuation, inhomogeneous appearance, and
inconsistent
texture. Subsequently, a region-dependent boundary refinement uses a
discriminative appearance classifier to refine the surface, and finally a
region-driven level set refinement is used to extract the fine scale surface
details.
Michal Sofka did his undergraduate work at the Czech Technical University. He
received the MS degree in Electrical Engineering from Union College in 2001. He
received the MS and PhD degrees in Computer Science from the Rensselaer
Polytechnic Institute (RPI) in 2006 and 2008, respectively. In 2004, he was a
technical employee at Siemens Corporate Research. He joined the same company as
a full time Research Scientist in 2008 and became a Project Manager in 2011.
Dr.
Sofka currently manages research and development projects for various Siemens
business units and external customers. His interests include machine learning,
segmentation, object detection, registration, and matching with the application
in medical imaging.
images
is difficult to obtain in most practical situations due to the large parameter
space and due to data variations. We address this problem by a Hierarchical
Detection Network (HDN) that detects the structures sequentially, one-by-one.
The interdependence of structures poses and prior information embedded in our
domain of medical images results in better performance than detecting the
structures individually. The posterior distribution of each structure pose is
approximated at every step by sequential Monte Carlo sampling. The samples are
propagated within the sequence across multiple objects and hierarchical levels.
The most reliable samples are used to initialize a multi-organ segmentation
system. An example of this procedure will be shown on segmenting pathological
lung in CT images. The lung pose initializes the detection of a set of
automatically-selected landmarks which typically lie near geometric structures
both within and outside the lung (e.g. ribs and spine). These stable landmarks
are used to robustly initialize a coarse statistical model of the lung shape
and
help to cope with higher attenuation, inhomogeneous appearance, and
inconsistent
texture. Subsequently, a region-dependent boundary refinement uses a
discriminative appearance classifier to refine the surface, and finally a
region-driven level set refinement is used to extract the fine scale surface
details.
Michal Sofka did his undergraduate work at the Czech Technical University. He
received the MS degree in Electrical Engineering from Union College in 2001. He
received the MS and PhD degrees in Computer Science from the Rensselaer
Polytechnic Institute (RPI) in 2006 and 2008, respectively. In 2004, he was a
technical employee at Siemens Corporate Research. He joined the same company as
a full time Research Scientist in 2008 and became a Project Manager in 2011.
Dr.
Sofka currently manages research and development projects for various Siemens
business units and external customers. His interests include machine learning,
segmentation, object detection, registration, and matching with the application
in medical imaging.