- Email: [email protected]
Automatic male pelvis segmentation from CT images via statistically trained multi-object deformable m-rep models
Proceedings of the 46th Annual ASTRO Meeting
Materials/Methods: In previous work, we have shown the limitations of fast spiral computed tomography (CT) in defining the true shape of a moving tumour. PET imaging of a moving sphere in air can provide accurate volumes encompassing tumour motion. In this work, an anthropomorphic phantom containing positron emitter distributed to simulate physiological concentrations of FDG was used. Two spheres (13 mm and 29 mm diameter) were used to simulate tumours. Each was oscillated within the lung of the phantom, with motion extents of 0, 7, 15, and 25 mm in each of the three principle planes. For each scenario, three different imaging modalities were used; spiral CT-simulation, PET, and digital fluoroscopy. The latter was assumed to be the gold standard. A CT-based gross tumour volume (GTV) was generated using a threshold of -875 HU. A population-based, symmetric margin of 15 mm, reflecting both motion and set-up uncertainties, was then added as per clinical practice to generate a CT-based PTV. A PET-based ITV was defined using a region-growing technique, starting with the voxel with the maximum value and extending to all voxels connected to the region containing a value greater than two standard deviations above lung background. A symmetric set-up margin of 7.5 mm was added to PET-based ITVs to create PTVs. This methodology was clinically validated using six patients with well-defined tumours. All image modalities were acquired on the same day and co-registered. Co-registered fluoroscopic movies were used to determine the extent of motion in the three principle planes. Results: For the physiological phantom studies, the PTVs based on CT and population-based margins were twice the volume as the PET-based PTVs in all cases. Figure 1 shows all PTVs plotted against the ideal PTV of the known volume. In all cases, the PET-based PTVs were closer to the corresponding ideal PTV than those based on CT. In no case would the PET-based PTV have resulted in a geographic miss. In the patient validation study, it was not always possible to accurately determine motion extents in all three directions using fluoroscopy. For those patients, where motion extents could be accurately assessed, both the CT-based PTV and the PET-based ITV encompassed the fluoroscopic motion, within experimental error. Conclusions: Experiments using a physiological torso phantom have shown that quantitatively segmented PET images can provide an accurate individualized ITV for moving lung tumours. Phantom results have been clinically validated in a group of lung cancer patients using integrated digital fluoroscopy. While PET may improve the localization of GTV, these results suggest that PET can define more optimal PTVs by accounting for a tumour’s unique motion pattern.
41
Automatic Male Pelvis Segmentation from CT Images via Statistically Trained Multi-Object Deformable M-rep Models
E. L. Chaney, S. Pizer, S. Joshi, R. Broadhurst, T. Fletcher, G. Gash, Q. Han, J. Jeong, C. Lu, D. Merck, J. Stough, G. Tracton Medical Image Display and Analysis Group, University of North Carolina, Chapel Hill, NC Purpose/Objective: The purpose of this study was to evaluate m-rep deformable models for automatic segmentation of multi-object soft-tissue complexes in low contrast images with application to bladder, rectum and prostate in planning and treatment CT images. M-rep deformation is guided by knowledge of object and inter-object geometry and image intensity patterns gained through a statistical training procedure. The method is broadly applicable but is tested here for adaptive radiotherapy for prostate cancer, where the segmentation is applied to CT images acquired in the treatment room over the course of radiotherapy. Materials/Methods: M-reps* are deformable models that describe anatomical objects in terms of a mean shape and local broadening & elongation, twisting & bending, and displacement of volume elements of the object(s) forming the mean. M-reps also model non-rigid between-object and within-object geometry such as the relationship between adjacent non-interpenetrating structures. Principal modes of shape variation computed in non-Euclidean space provide statistics of variability of the object(s) within a training population. Separately, image intensity statistics are generated in m-rep object-relative coordinates from the same training images. The mean model is deformed in a target image by successive optimization to yield the most probable segmentation given the image data. The objective function reflects both the shape statistics and the intensity statistics produced at training. The deformation proceeds from the multi-object complex as a whole, to single objects starting from inter-object predictions, to intra-object volume sections. For the male pelvis the supra-pubic bones are segmented first, followed by the bladder/rectum/ prostate complex, and finally the individual soft tissue structures. In this study of the applicability of m-reps for adaptive therapy, the same-patient “mean” model and image intensity training
S153
S154
I. J. Radiation Oncology
● Biology ● Physics
Volume 60, Number 1, Supplement, 2004
were derived from the planning image. Also the shape statistics in this study describe the transformation between the planning image model and the treatment images. Results: M-rep segmentations obtained in this study were compared to expert human segmentations by statistical analysis of metrics including volume overlap, mean surface separation, maximum surface separation, and separation-distance histogram quartiles. Clinically acceptable, non-interpenetrating segmentations of the prostate and other soft tissues were obtained. Conclusions: Automatic m-rep segmentation works well for adaptive radiotherapy for prostate cancer. The effectiveness of m-reps is attributed to several important properties: 1) the ability to represent solid volume geometry; 2) geometric modeling that reflects both local shape and local inter-regional relations; 3) the inclusion of statistics describing geometric properties of individual objects and object groups and of inter-relationships between these objects; 4) the inclusion of statistical aspects of intensity patterns in object-relative geometry; and 5) the multiscale nature both of training and of segmentation of target images. The next step is generalization to the multi-patient situation. This will involve more stages of statistical training, more stages in the model deformation process, and greater attention to the variability of relationships between the various structures comprising the multi-object m-rep model. Acknowledgements: J. Bechtel, MD and J. Rosenman, MD for image contouring, and Y-Y. Chi, BS and K. Muller, PhD for statistical analysis of data. Research supported by NIH P01 EB002779 *S Pizer et al, Deformable M-Reps for 3D Medical Image Segmentation, Int J Comp Vision, 55: 85–106 (2003)
42
Comparison of Biological Models to Predict the Incidence of Breast Radiotherapy-induced Cardiac Perfusion Defects
S. K. Das,1 A. Baydush,1 S. Zhou,1 M. Miften,1 X. Yu,1 K. Light,1 T. Wong,1 M. Blazing,2 L. Marks1 1 Radiation Oncology, Duke University, Durham, NC, 2Cardiology, Duke University, Durham, NC Purpose/Objective: Predicting the risk of radiotherapy (RT) induced cardiac toxicity plays an important role in RT field design. Commonly used predictive models assume that the dose-response behavior can be characterized by pre-defined mathematical functions (parametric models), even though these function forms are speculative and hence may not reflect clinical reality. The subject of this work is to determine if models that do not assume any such pre-defined mathematical behavior (non-parametric models) can be better predictors of left breast RT-induced cardiac toxicity. Materials/Methods: Since 1998, we have been conducting an IRB-approved prospective clinical trial in left-sided breast/chest wall cancer patients receiving tangential photon therapy (46 –50 Gy at 1.8 –2 Gy/fx) to assess RT-induced changes in left ventricular (LV) function. LV functional changes were scored by single photon emission computed tomography (SPECT) perfusion defects. The basis of this study is 68 patients with normal pre-RT SPECT scans. Of these, 19 developed defects at 6 months post-RT. The dose-volume histograms (DVH) of all patients were used as inputs to a non-parametric linear discriminant analysis model (LDA) and three commonly used parametric models: Lyman normal tissue complication probability (LNTCP), relative seriality (RS), and generalized equivalent uniform dose (GEUD). The pre-defined mathematical functions of the parametric models were fitted to the data using statistical methods (maximum likelihood estimation and F-test). The non-parametric LDA model selected a linear combination of volumes above certain dose levels to best separate the groups with and without defects. The models were compared to each other using receiver operating characteristics (ROC) curves, which plot sensitivity vs. 1-specificity (greater area under the ROC curve implies a more accurate model). Optimistic estimates of each model’s predictive capabilities were obtained by generating ROC curves using all patient datasets in the model generation and error estimation (train-test-all). Additionally, pessimistic estimates were obtained using a technique (leave-one-out) where each patient dataset is used as a validation set with the remainder as the model generation set. Results: ROC areas under the curves for train-test-all (leave-one-out) were 0.80 (0.75) for NTCP, 0.80 (0.78) for RS, 0.80 (0.74) for GEUD, and 0.91 (0.87) for LDA (see figure). Among the parametric models (NTCP, RS, GEUD), a test of the superiority of one over the other was not significant (p ⫽ 0.44), which can be qualitatively seen in the similarity of their ROC curves. The non-parametric LDA model was a significantly better predictor than the parametric models in the optimistic estimate (p ⫽ 0.03), but less so in the pessimistic estimate (p ⫽ 0.11). LDA selected volumes above 30 and 58 Gy as most important in separating the groups. Conclusions: As a predictor of RT-induced left-ventricular perfusion defects, commonly used parametric predictive models (which assume that the dose-response behavior is characterized by pre-defined mathematical functions) appear to be inferior to the linear discriminant analysis model (which does not assume a dose-response behavior). Linear discriminant analysis essentially identified two left-ventricular volumes as most important in separating the groups: volume within the fields, and volume at high dose. (Acknowledgements: DOD grants DAMD 17-98-1-8071 and BC 010663; PLUNC, Univ. N. Carolina).
Materials/Methods: In previous work, we have shown the limitations of fast spiral computed tomography (CT) in defining the true shape of a moving tumour. PET imaging of a moving sphere in air can provide accurate volumes encompassing tumour motion. In this work, an anthropomorphic phantom containing positron emitter distributed to simulate physiological concentrations of FDG was used. Two spheres (13 mm and 29 mm diameter) were used to simulate tumours. Each was oscillated within the lung of the phantom, with motion extents of 0, 7, 15, and 25 mm in each of the three principle planes. For each scenario, three different imaging modalities were used; spiral CT-simulation, PET, and digital fluoroscopy. The latter was assumed to be the gold standard. A CT-based gross tumour volume (GTV) was generated using a threshold of -875 HU. A population-based, symmetric margin of 15 mm, reflecting both motion and set-up uncertainties, was then added as per clinical practice to generate a CT-based PTV. A PET-based ITV was defined using a region-growing technique, starting with the voxel with the maximum value and extending to all voxels connected to the region containing a value greater than two standard deviations above lung background. A symmetric set-up margin of 7.5 mm was added to PET-based ITVs to create PTVs. This methodology was clinically validated using six patients with well-defined tumours. All image modalities were acquired on the same day and co-registered. Co-registered fluoroscopic movies were used to determine the extent of motion in the three principle planes. Results: For the physiological phantom studies, the PTVs based on CT and population-based margins were twice the volume as the PET-based PTVs in all cases. Figure 1 shows all PTVs plotted against the ideal PTV of the known volume. In all cases, the PET-based PTVs were closer to the corresponding ideal PTV than those based on CT. In no case would the PET-based PTV have resulted in a geographic miss. In the patient validation study, it was not always possible to accurately determine motion extents in all three directions using fluoroscopy. For those patients, where motion extents could be accurately assessed, both the CT-based PTV and the PET-based ITV encompassed the fluoroscopic motion, within experimental error. Conclusions: Experiments using a physiological torso phantom have shown that quantitatively segmented PET images can provide an accurate individualized ITV for moving lung tumours. Phantom results have been clinically validated in a group of lung cancer patients using integrated digital fluoroscopy. While PET may improve the localization of GTV, these results suggest that PET can define more optimal PTVs by accounting for a tumour’s unique motion pattern.
41
Automatic Male Pelvis Segmentation from CT Images via Statistically Trained Multi-Object Deformable M-rep Models
E. L. Chaney, S. Pizer, S. Joshi, R. Broadhurst, T. Fletcher, G. Gash, Q. Han, J. Jeong, C. Lu, D. Merck, J. Stough, G. Tracton Medical Image Display and Analysis Group, University of North Carolina, Chapel Hill, NC Purpose/Objective: The purpose of this study was to evaluate m-rep deformable models for automatic segmentation of multi-object soft-tissue complexes in low contrast images with application to bladder, rectum and prostate in planning and treatment CT images. M-rep deformation is guided by knowledge of object and inter-object geometry and image intensity patterns gained through a statistical training procedure. The method is broadly applicable but is tested here for adaptive radiotherapy for prostate cancer, where the segmentation is applied to CT images acquired in the treatment room over the course of radiotherapy. Materials/Methods: M-reps* are deformable models that describe anatomical objects in terms of a mean shape and local broadening & elongation, twisting & bending, and displacement of volume elements of the object(s) forming the mean. M-reps also model non-rigid between-object and within-object geometry such as the relationship between adjacent non-interpenetrating structures. Principal modes of shape variation computed in non-Euclidean space provide statistics of variability of the object(s) within a training population. Separately, image intensity statistics are generated in m-rep object-relative coordinates from the same training images. The mean model is deformed in a target image by successive optimization to yield the most probable segmentation given the image data. The objective function reflects both the shape statistics and the intensity statistics produced at training. The deformation proceeds from the multi-object complex as a whole, to single objects starting from inter-object predictions, to intra-object volume sections. For the male pelvis the supra-pubic bones are segmented first, followed by the bladder/rectum/ prostate complex, and finally the individual soft tissue structures. In this study of the applicability of m-reps for adaptive therapy, the same-patient “mean” model and image intensity training
S153
S154
I. J. Radiation Oncology
● Biology ● Physics
Volume 60, Number 1, Supplement, 2004
were derived from the planning image. Also the shape statistics in this study describe the transformation between the planning image model and the treatment images. Results: M-rep segmentations obtained in this study were compared to expert human segmentations by statistical analysis of metrics including volume overlap, mean surface separation, maximum surface separation, and separation-distance histogram quartiles. Clinically acceptable, non-interpenetrating segmentations of the prostate and other soft tissues were obtained. Conclusions: Automatic m-rep segmentation works well for adaptive radiotherapy for prostate cancer. The effectiveness of m-reps is attributed to several important properties: 1) the ability to represent solid volume geometry; 2) geometric modeling that reflects both local shape and local inter-regional relations; 3) the inclusion of statistics describing geometric properties of individual objects and object groups and of inter-relationships between these objects; 4) the inclusion of statistical aspects of intensity patterns in object-relative geometry; and 5) the multiscale nature both of training and of segmentation of target images. The next step is generalization to the multi-patient situation. This will involve more stages of statistical training, more stages in the model deformation process, and greater attention to the variability of relationships between the various structures comprising the multi-object m-rep model. Acknowledgements: J. Bechtel, MD and J. Rosenman, MD for image contouring, and Y-Y. Chi, BS and K. Muller, PhD for statistical analysis of data. Research supported by NIH P01 EB002779 *S Pizer et al, Deformable M-Reps for 3D Medical Image Segmentation, Int J Comp Vision, 55: 85–106 (2003)
42
Comparison of Biological Models to Predict the Incidence of Breast Radiotherapy-induced Cardiac Perfusion Defects
S. K. Das,1 A. Baydush,1 S. Zhou,1 M. Miften,1 X. Yu,1 K. Light,1 T. Wong,1 M. Blazing,2 L. Marks1 1 Radiation Oncology, Duke University, Durham, NC, 2Cardiology, Duke University, Durham, NC Purpose/Objective: Predicting the risk of radiotherapy (RT) induced cardiac toxicity plays an important role in RT field design. Commonly used predictive models assume that the dose-response behavior can be characterized by pre-defined mathematical functions (parametric models), even though these function forms are speculative and hence may not reflect clinical reality. The subject of this work is to determine if models that do not assume any such pre-defined mathematical behavior (non-parametric models) can be better predictors of left breast RT-induced cardiac toxicity. Materials/Methods: Since 1998, we have been conducting an IRB-approved prospective clinical trial in left-sided breast/chest wall cancer patients receiving tangential photon therapy (46 –50 Gy at 1.8 –2 Gy/fx) to assess RT-induced changes in left ventricular (LV) function. LV functional changes were scored by single photon emission computed tomography (SPECT) perfusion defects. The basis of this study is 68 patients with normal pre-RT SPECT scans. Of these, 19 developed defects at 6 months post-RT. The dose-volume histograms (DVH) of all patients were used as inputs to a non-parametric linear discriminant analysis model (LDA) and three commonly used parametric models: Lyman normal tissue complication probability (LNTCP), relative seriality (RS), and generalized equivalent uniform dose (GEUD). The pre-defined mathematical functions of the parametric models were fitted to the data using statistical methods (maximum likelihood estimation and F-test). The non-parametric LDA model selected a linear combination of volumes above certain dose levels to best separate the groups with and without defects. The models were compared to each other using receiver operating characteristics (ROC) curves, which plot sensitivity vs. 1-specificity (greater area under the ROC curve implies a more accurate model). Optimistic estimates of each model’s predictive capabilities were obtained by generating ROC curves using all patient datasets in the model generation and error estimation (train-test-all). Additionally, pessimistic estimates were obtained using a technique (leave-one-out) where each patient dataset is used as a validation set with the remainder as the model generation set. Results: ROC areas under the curves for train-test-all (leave-one-out) were 0.80 (0.75) for NTCP, 0.80 (0.78) for RS, 0.80 (0.74) for GEUD, and 0.91 (0.87) for LDA (see figure). Among the parametric models (NTCP, RS, GEUD), a test of the superiority of one over the other was not significant (p ⫽ 0.44), which can be qualitatively seen in the similarity of their ROC curves. The non-parametric LDA model was a significantly better predictor than the parametric models in the optimistic estimate (p ⫽ 0.03), but less so in the pessimistic estimate (p ⫽ 0.11). LDA selected volumes above 30 and 58 Gy as most important in separating the groups. Conclusions: As a predictor of RT-induced left-ventricular perfusion defects, commonly used parametric predictive models (which assume that the dose-response behavior is characterized by pre-defined mathematical functions) appear to be inferior to the linear discriminant analysis model (which does not assume a dose-response behavior). Linear discriminant analysis essentially identified two left-ventricular volumes as most important in separating the groups: volume within the fields, and volume at high dose. (Acknowledgements: DOD grants DAMD 17-98-1-8071 and BC 010663; PLUNC, Univ. N. Carolina).