Automatic segmentation of SPECT/CT-images using a deformable surface model based on Fourier descriptors with application in radionuclide therapy
(2011)Medical Physics Programme
- Abstract (Swedish)
- Introduction: The purpose of this study was to develop a technique for automatic segmentation based on Fourier descriptors for separating smooth objects from the background in SPECT-images. The aim of the segmentation method was to be able to automatically outline kidneys in SPECT-images of patients undergoing radionuclide therapy with 177Lu-DOTATATE. The potential advantage of combining SPECT- and CT-images for the task of segmentation using the same technique was also explored.
Material and methods: A method for segmentation of SPECT- and SPECT/CT-images was developed and implemented in IDL. The process included a first crude estimation of the organ position and size in a graphical user interface, producing an ellipsoid for... (More) - Introduction: The purpose of this study was to develop a technique for automatic segmentation based on Fourier descriptors for separating smooth objects from the background in SPECT-images. The aim of the segmentation method was to be able to automatically outline kidneys in SPECT-images of patients undergoing radionuclide therapy with 177Lu-DOTATATE. The potential advantage of combining SPECT- and CT-images for the task of segmentation using the same technique was also explored.
Material and methods: A method for segmentation of SPECT- and SPECT/CT-images was developed and implemented in IDL. The process included a first crude estimation of the organ position and size in a graphical user interface, producing an ellipsoid for initialisation of the coming optimisation procedure. The ellipsoid was used as input to an optimisation process where a surface, described by Fourier descriptors of successively increasing order, was adapted to a boundary measure based on the grey level gradient of the image. In order to avoid uncontrolled growth of the surface, the objective function was normalised to the total surface area raised to an exponent, termed the normalisation exponent. The surface description of the object could finally be transformed into a volume description of the object, i.e. a labelling of the voxels encompassed by the surface, by using a method based on the relative directions of the surface normal and the directed line segment joining a voxel coordinate and its closest point on the surface. The performance of the method was evaluated in different test images. Single organ images based on the left kidney of the XCAT anthropomorphic computer phantom were used in a first step, with different contrast to noise ratios and spatial resolutions. The degree of complexity was then increased by using test images designed to imitate the biodistribution of 177Lu-DOTATATE in patients undergoing radionuclide therapy, again using the XCAT phantom. The same distribution was also used to create Monte Carlo simulated SPECT-images, to also model the actual camera characteristics. Finally the segmentation was performed in a set of real clinical images. In the simulated images the performance of the method was quantified both in terms of the distance between the estimated surface and the true surface, and in terms of a volume measure. In the single organ images evaluation was performed as a function of the number of Fourier orders used in the segmentation, for varying contrast to noise ratios and varying spatial resolution. Also investigated was the behaviour with respect to the normalisation exponent. The sensitivity of the method to a heterogeneous background distribution was performed in the simulated images mimicking the biodistribution of 177Lu-DOTATATE. Finally the operator dependence, i.e. the initialisation dependence, on the final segmentation result was tested in the set of clinical images.
Result: The segmentation method was generally found to be resistant to noise. For a contrast to noise ratio of approximately unity combined with a high enough spatial resolution, the mean distance between the segmented surface and a reference surface was below one voxel side. With a poorer spatial resolution the result deteriorated for contrast to noise ratios of approximately one to two. When a background was added to the images, imitating a clinical activity distribution, the segmentation method showed some sensitivity to nearby high signal objects, for instance for the spleen which in some patients is in close to the left kidney. The operator dependence of the method when applied to clinical images was found to be low. No obvious advantage was observed when CT- images were combined with SPECT-images as the basis for the segmentation.
Discussion: The segmentation method gives a good agreement with the original volume segmented and has the potential of being useful as a tool for clinical SPECT-images. It is insensitive to noise, and is particularly suited for smooth images, such as those obtained with SPECT imaging. The inclusion of CT-images was found not to contribute to the precision of the method. A possible explanation is that the kidneys are in themselves smooth objects and no basic information of the object characteristics is added by the CT-images. Another explanation may be that the weighting used of the SPECT- and CT-based information was suboptimal. In cases of nearby high signal objects close to the object of interest, the method showed a tendency of incorporating the nearby object. One way to handle this problem could be to incorporate a priori information into the segmentation model, of interest for future work. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/2273034
- author
- Gustavsson, Johan
- supervisor
- organization
- year
- 2011
- type
- H2 - Master's Degree (Two Years)
- subject
- keywords
- nuklearmedicin
- language
- English
- id
- 2273034
- date added to LUP
- 2012-01-02 15:18:07
- date last changed
- 2012-01-02 15:18:07
@misc{2273034,
abstract = {{Introduction: The purpose of this study was to develop a technique for automatic segmentation based on Fourier descriptors for separating smooth objects from the background in SPECT-images. The aim of the segmentation method was to be able to automatically outline kidneys in SPECT-images of patients undergoing radionuclide therapy with 177Lu-DOTATATE. The potential advantage of combining SPECT- and CT-images for the task of segmentation using the same technique was also explored.
Material and methods: A method for segmentation of SPECT- and SPECT/CT-images was developed and implemented in IDL. The process included a first crude estimation of the organ position and size in a graphical user interface, producing an ellipsoid for initialisation of the coming optimisation procedure. The ellipsoid was used as input to an optimisation process where a surface, described by Fourier descriptors of successively increasing order, was adapted to a boundary measure based on the grey level gradient of the image. In order to avoid uncontrolled growth of the surface, the objective function was normalised to the total surface area raised to an exponent, termed the normalisation exponent. The surface description of the object could finally be transformed into a volume description of the object, i.e. a labelling of the voxels encompassed by the surface, by using a method based on the relative directions of the surface normal and the directed line segment joining a voxel coordinate and its closest point on the surface. The performance of the method was evaluated in different test images. Single organ images based on the left kidney of the XCAT anthropomorphic computer phantom were used in a first step, with different contrast to noise ratios and spatial resolutions. The degree of complexity was then increased by using test images designed to imitate the biodistribution of 177Lu-DOTATATE in patients undergoing radionuclide therapy, again using the XCAT phantom. The same distribution was also used to create Monte Carlo simulated SPECT-images, to also model the actual camera characteristics. Finally the segmentation was performed in a set of real clinical images. In the simulated images the performance of the method was quantified both in terms of the distance between the estimated surface and the true surface, and in terms of a volume measure. In the single organ images evaluation was performed as a function of the number of Fourier orders used in the segmentation, for varying contrast to noise ratios and varying spatial resolution. Also investigated was the behaviour with respect to the normalisation exponent. The sensitivity of the method to a heterogeneous background distribution was performed in the simulated images mimicking the biodistribution of 177Lu-DOTATATE. Finally the operator dependence, i.e. the initialisation dependence, on the final segmentation result was tested in the set of clinical images.
Result: The segmentation method was generally found to be resistant to noise. For a contrast to noise ratio of approximately unity combined with a high enough spatial resolution, the mean distance between the segmented surface and a reference surface was below one voxel side. With a poorer spatial resolution the result deteriorated for contrast to noise ratios of approximately one to two. When a background was added to the images, imitating a clinical activity distribution, the segmentation method showed some sensitivity to nearby high signal objects, for instance for the spleen which in some patients is in close to the left kidney. The operator dependence of the method when applied to clinical images was found to be low. No obvious advantage was observed when CT- images were combined with SPECT-images as the basis for the segmentation.
Discussion: The segmentation method gives a good agreement with the original volume segmented and has the potential of being useful as a tool for clinical SPECT-images. It is insensitive to noise, and is particularly suited for smooth images, such as those obtained with SPECT imaging. The inclusion of CT-images was found not to contribute to the precision of the method. A possible explanation is that the kidneys are in themselves smooth objects and no basic information of the object characteristics is added by the CT-images. Another explanation may be that the weighting used of the SPECT- and CT-based information was suboptimal. In cases of nearby high signal objects close to the object of interest, the method showed a tendency of incorporating the nearby object. One way to handle this problem could be to incorporate a priori information into the segmentation model, of interest for future work.}},
author = {{Gustavsson, Johan}},
language = {{eng}},
note = {{Student Paper}},
title = {{Automatic segmentation of SPECT/CT-images using a deformable surface model based on Fourier descriptors with application in radionuclide therapy}},
year = {{2011}},
}