Functional nuclear imaging provides unique information that is complementary to structural imaging modalities like CT or MR which provide a view of the patient's anatomy. Functional imaging modalities like SPECT or PET on the other hand provide a view into physiological activities, whether that is metabolic function, blood flow or the chemical composition of tissues. This information is widely used for diagnosis, but also very valuable in the interventional case.
Time constraints may make the application of computationally expensive algorithms unfeasible in clinical settings. This is true in diagnostic settings, but the time constraints are even stricter for intra-operative purposes. The tightly orchestrated workflow in the operating room (often involving a large support staff) does not allow for costly down times, but the information provided during surgery becomes quickly inaccurate or even invalid because of tissue manipulation at the same time. So while time is constrained, there is a constant need for updated information.
This thesis develops accelerated methods for the registration and reconstruction of functional nuclear volumes, making the application of these algorithms feasible in clinical practice. The first chapters develop the fundamental principles of registration and algorithms for massively parallel hardware that are expanded upon in later chapters for the development of the methods introduced in this thesis.
In order to achieve the required interactive speed, hardware-based acceleration of existing algorithms is one option, the introduction of novel algorithms that replace slower ones the second option. This dissertation demonstrates both approaches.
The contributions of this dissertation include one of the first registration pipelines for deformable registration fully accelerated by graphics processing units (GPUs). Using statistical intensity priors, this method allows the accurate and fast registration between different modalities. The next contribution is the careful optimization of the main bottleneck for registration methods using statistical intensity priors on massively parallel hardware architectures like GPUs: The optimization of the joint histogram computation with its parallelization unfriendly memory access patterns. The third contribution is a GPU-accelerated pipeline for SPECT reconstruction.
The main contribution of this thesis lies in a new registration method between a 3D volume (acquired pre- or intraoperatively) and intra-operative tracked probe measurements for radio-guided surgery. The registration allows rapid updates of the image guidance for the surgeon, while only requiring little time. The usefulness of this approach is demonstrated for the case of intra-operative freehand SPECT reconstruction during sentinel lymph node biopsies for breast cancer treatment. Using a previous 3D volume as prior knowledge, the 1D-3D registration allows the fast update of reconstruction volumes by performing a registration to 1D measurements provided by a gamma probe. This replaces in effect a reconstruction step with registration using the prior knowledge. Finally, we provide an outlook towards further research necessary in order to make this a viable alternative to current techniques.
Functional nuclear imaging provides unique information that is complementary to structural imaging modalities like CT or MR which provide a view of the patient's anatomy. Functional imaging modalities like SPECT or PET on the other hand provide a view into physiological activities, whether that is metabolic function, blood flow or the chemical composition of tissues. This information is widely used for diagnosis, but also very valuable in the interventional case.
Time constraints may make the application of computationally expensive algorithms unfeasible in clinical settings. This is true in diagnostic settings, but the time constraints are even stricter for intra-operative purposes. The tightly orchestrated workflow in the operating room (often involving a large support staff) does not allow for costly down times, but the information provided during surgery becomes quickly inaccurate or even invalid because of tissue manipulation at the same time. So while time is constrained, there is a constant need for updated information.
This thesis develops accelerated methods for the registration and reconstruction of functional nuclear volumes, making the application of these algorithms feasible in clinical practice. The first chapters develop the fundamental principles of registration and algorithms for massively parallel hardware that are expanded upon in later chapters for the development of the methods introduced in this thesis.
In order to achieve the required interactive speed, hardware-based acceleration of existing algorithms is one option, the introduction of novel algorithms that replace slower ones the second option. This dissertation demonstrates both approaches.
The contributions of this dissertation include one of the first registration pipelines for deformable registration fully accelerated by graphics processing units (GPUs). Using statistical intensity priors, this method allows the accurate and fast registration between different modalities. The next contribution is the careful optimization of the main bottleneck for registration methods using statistical intensity priors on massively parallel hardware architectures like GPUs: The optimization of the joint histogram computation with its parallelization unfriendly memory access patterns. The third contribution is a GPU-accelerated pipeline for SPECT reconstruction.
The main contribution of this thesis lies in a new registration method between a 3D volume (acquired pre- or intraoperatively) and intra-operative tracked probe measurements for radio-guided surgery. The registration allows rapid updates of the image guidance for the surgeon, while only requiring little time. The usefulness of this approach is demonstrated for the case of intra-operative freehand SPECT reconstruction during sentinel lymph node biopsies for breast cancer treatment. Using a previous 3D volume as prior knowledge, the 1D-3D registration allows the fast update of reconstruction volumes by performing a registration to 1D measurements provided by a gamma probe. This replaces in effect a reconstruction step with registration using the prior knowledge. Finally, we provide an outlook towards further research necessary in order to make this a viable alternative to current techniques.
«
Functional nuclear imaging provides unique information that is complementary to structural imaging modalities like CT or MR which provide a view of the patient's anatomy. Functional imaging modalities like SPECT or PET on the other hand provide a view into physiological activities, whether that is metabolic function, blood flow or the chemical composition of tissues. This information is widely used for diagnosis, but also very valuable in the interventional case.
Time constraints may make the a...
»