Automated Delineation of Target Regions for Radiation Treatment Planning
The GTV and the clinical target volume (CTV) (larger than the GTV, the CTV defines the volume including gross tumor and areas of likely microscopic involvement) are the starting points in radiation treatment planning as mentioned above. There are two reasons why computer-assisted delineation is important:
- (1) The intra- and interobserver variability of GTV delineation by radiation oncologists is plagued by low reproducibility [127-130]. Chao et al. reported that the variation in contours in CTVs delineated by radiation oncologists is a critical issue and that using computer-assisted methods reduced volumetric variation and improved geometric consistency in cases of base of tongue and nasopharyngeal cancer .
- (2) Manual contouring is tedious and time-consuming. According to one study, the average time saved by using a computer-assisted method is 26% to 29% for experienced physicians and 38% to 47% for less-experienced physicians .
Automated contouring based on statistical CTV models is one of the approaches to reduce the inter- and intraobserver variability in GTV or CTV contour generation. Arimura developed a computational method for producing statistical CTV shape models for low-, intermediate-, and high-risk prostate cancers based on a point distribution model, which can be used as a CTV template for automated contouring in prostate cancer radiation treatment planning . First, fifteen radiation oncologists delineated CTV contours for three risk levels. The low-risk, intermediate-risk, and high-risk CTVs included the prostate, the prostate plus the proximal 1 cm of the seminal vesicles, and the prostate plus the proximal 2 cm of the seminal vesicles, respectively. The statistical CTV models for the three risk types were derived based on principal component analysis (PCA), which statistically incorporated the interobserver variability. CTV surfaces were triangulated using a marching cubes algorithm. For matching the number of points on the surfaces of all CTV regions, the number of vertices on each CTV polygonal surface was reduced to 1,000 with quadric error metrics. All CTV regions were registered with a reference CTV using an iterative closest point algorithm for calculation of a covariance matrix to be employed for the PCA-based CTV modeling. CTV models of the three risk types were produced, which consisted of a mean CTV and PCA coefficients multiplied by eigenvectors. Figure 4.23 shows shape variations of statistical CTV models
Fig. 4.23 Shape variations of statistical CTV models produced by the first and second largest modes for an intermediate-risk group produced by the first and second largest modes for an intermediate-risk group. These computational anatomical techniques and mathematical modeling of targets and organs may allow for adaptive target delineation during the course of radiotherapy.
A number of automated delineation methods for determining the GTV or CTV have been proposed to reduce the inter- and intraobserver variability and planning time, as well as to increase the segmentation accuracy of the GTV. Methods using PET images are based on thresholding of the standardized uptake value (SUV) [132, 133]; region-growing methods also use the SUV . Gaussian mixture model- based segmentation , gradient-based segmentation methods , the fuzzy locally adaptive Bayesian approach , the fuzzy c-means algorithm , and model-based methods  have all been studied. Methods such as MR atlas-based  and probabilistic atlas-based  segmentation have also been proposed.
Segmentation methods for GTV based on positron-emission tomogra- phy/computed tomography (PET/CT) datasets, which include metabolic as well as morphological information, have been assessed. A tumor’s higher rate of aerobic glycolysis is directly quantified by 18F-fluorodeoxyglucose (FDG) PET. El Naqa et al. developed a multimodality segmentation method using a multivalued level set method by combining imaging data obtained from different modalities, including PET/CT . In their study, the level set method was applied to a vector image including CT and PET images so that an energy function could be minimized for determination of CTV regions. As a result, the corresponding Dice similarity coefficient was 0.90 ± 0.02 when CT, MR, and PET images were used. We  attempted to incorporate PET biological and CT morphological information on tumor contours determined by radiation oncologists into an optimum contour selection (OCS) framework  using a machine learning protocol. We have proposed an automated method for extracting GTVs using a machine learning classifier that accumulates radiation oncology datasets of planning CT and FDG- PET/CT images.
Our method [143, 144] is to feed GTV contours determined by radiation oncologists into a machine learning classifier during the training step, after which the classifier produces the “degree of GTV” for each voxel in the testing step. Six voxel-based image features, including voxel values and magnitudes of image gradient vectors, are derived from each voxel using the planning CT and PET/CT image datasets. Initially, lung tumors are extracted using a support vector machine (SVM) that learns six voxel-based features inside and outside each tumor region determined by radiation oncologists. The final tumor regions are determined using the OCS approach that can be used for selection of a global optimum object contour based on multiple delineations with a level set method around the tumor. Figure 4.24 shows an SVM output image and GTV contours that were multiply delineated using the proposed method on the planning CT image at evolution times of 0, 2000, 3337 (optimum contour), 5000, and 6000. The proposed method achieved an average Dice similarity coefficient of 0.84 in six lung cancer patients, while the conventional method output was 0.78.
K. Mori et al.
Fig. 4.24 SVM output image and GTV contours that were multiply delineated using the proposed method on the planning CT image at evolution times of 0, 2000, 3337 (optimum contour), 5000, and 6000
Similar-Case-Based Beam Arrangements in Stereotactic Body Radiotherapy
Stereotactic body radiotherapy (SBRT) aims to administer high enough doses to ablate a cancer while minimizing the dose to the surrounding healthy tissues by means of multiple beam arrangements, which generally consist of a large number of coplanar and noncoplanar beams . The determination of beam arrangements is time-consuming, and it is a demanding procedure for less-experienced treatment planners. Treatment planners’ skills are generally developed by repeated planning in clinical practice. The planners memorize many planning patterns as an evolving “database,” which can be searched for past cases similar to the case under consideration. Several studies have tried to develop computer-assisted methods for this modality [146-149]. Commowick et al. developed a method for selection of a template image, which is the most similar image selected with a distance between transformations in a radiation treatment planning database and can be employed in atlas-based segmentation . Chanyavanich et al. demonstrated the usefulness of prior treatment plans, derived from similar cases, to generate new intensity- modulated radiation therapy (IMRT) plans in prostate neoplasms . Mishra et al. developed a case-based reasoning system, which enables the use of knowledge and experience gained by oncologists in dealing with new patients . Schlaefer et al. reported the feasibility of a framework of case-based beam generation for robotic radiosurgery, which could reduce planning time while maintaining high plan quality for typical clinical cases with similar anatomy .
Magome et al. also developed a method for determination of beam arrangements based on similar cases, which proved to be helpful for making new plans in lung cancer SBRT . Beam arrangements were automatically determined based on similar cases using the following two steps: First, the five most similar cases to the current case were retrieved using geometrical features associated with the location, size, and shape of the planning target volume (PTV, the CTV plus a margin allowing for patient movement, position changes, and other variables), lung, and spinal cord. Then, five beam arrangements for the current case were automatically created by aligning five similar cases with the current case in terms of lung regions by use of a linear registration technique. To evaluate the beam arrangements, five treatment plans were manually designed by applying the beam arrangements to the current case. The most useful beam arrangement was chosen by sorting the five treatment plans based on several plan evaluation indices including the D95 (dose that covers 95% of the PTV), mean lung dose, and spinal cord maximum dose. They applied the proposed method to ten test cases by searching in an RTP database of 81 cases of lung cancer and compared the plan evaluation indices between the original treatment plan and the corresponding most useful similar- case-based treatment plan. The method had no statistically significant differences from the original beam arrangements (p > 0.05) with respect to the plan evaluation indices. This method could be employed as an educational tool for less-experienced treatment planners. Magome et al. developed a similar-case-based optimization method for beam arrangements in lung cancer SBRT for assisting treatment planners . The local beam direction optimization algorithm, which was developed in their study, improved the quality of treatment plans with significant differences (p < 0.05) in the homogeneity index and conformity index for the PTV, V10 (volume receiving >10 Gy), V20 (volume receiving dose >20 Gy), mean dose, and NTCP (normal tissue complication probability) for the lung.
The surrounding anatomical environments of tumors, which may affect RTP, were not considered in the study of . We developed a computational framework of retrieving similar cases using a local gradient distribution (LGD) feature for SBRT . We assumed that the LGD feature represents the surrounding anatomical environments of tumors. We adopted a local image descriptor, which was based on scale invariant feature transform . This proposed framework consists of two steps: searching and rearrangement. In the searching step, ten cases most similar to the current case are retrieved from the RTP database based on the shape similarity of a two-dimensional lung region at an isocenter plane. Next, the five most similar cases are selected using geometric features related to the location, size, and shape of the planning target volume, the lung, and the spinal cord. In the rearrangement step, the similarity rank of five selected cases is changed by use of the Euclidean distance between two LGD feature vectors. This is a similarity index based on the magnitudes and orientations of image gradients within a region of interest (ROI)
K. Mori et al.
around the isocenter. The gradient magnitude m(x, y/ and orientation 9(x, y/ are computed by the following equations based on a gradient vector (Sx(x, y/,Sy(x, y/) obtained by a Sobel filter:
The histogram of 36 discrete orientations is constructed by the following equation:
where the weighted magnitude w(x, y/ is obtained by multiplying the gradient magnitude by a Gaussian function with the scale estimated at the isocenter and 1 [в', 9(x, y/] is the Kronecker delta. The orientation with the highest value in the histogram Н(в'/ is considered the major orientation.
Gradient magnitudes and orientations of image gradient vectors are recalculated for each pixel in the ROI, which is divided into 4 x 4 subregions. Figure 4.25 shows illustrations of derivation of the LGD feature. The arrow length corresponds to the sum of the gradient magnitudes as shown in Fig. 4.25b.
An orientation histogram weighted by the vector magnitudes is produced in each subregion using eight bins covering 360°, as shown in Fig. 4.25b. The orientation histogram represents the relationship between eight orientations and the sum of
Fig. 4.25 Illustrations of derivation of the local image descriptor: (a) a CT image with an ROI and an arrow showing the major gradient orientation, (b) magnitudes of gradient vectors in 16 subregions, (c) a local image descriptor. The arrow length corresponds to the sum of the gradient magnitude in (b) gradient magnitudes by the following equation:
Finally, an LGD feature is assembled from an orientation histogram in 16 subregions, which is composed of 128 gradient magnitude features as shown in Fig. 4.25c. As a result, the cases, which are selected as cases similar to the test cases by the proposed method, resemble the test cases more than those selected by the method without the LGD features, in terms of the tumor location. This suggests that the use of the LGD feature is important in providing similar cases to treatment planners.
To evaluate Nonaka’s method , we applied the similar-case-based beam arrangement method, which was developed by Magome et al. . Figure 4.26 shows a plan generated using the original beam arrangement and five plans determined by similar-case-based beam arrangements, which were generated using the proposed method . The method has the potential to provide superior beam arrangements from the treatment planning point of view.
Quantitative Evaluation of the Robustness of Beam Directions for Charged Particle Therapy
The finely adjusted dose distribution produced in charged particle therapy such as proton or carbon ion beams is vulnerable to setup errors and/or organ motion. We investigated the quantification of the robustness of particle beam directions against patient setup errors in charged particle therapy . Power spectral analysis of target water-equivalent path length (WEPL) images in beam’s eye views was
Fig. 4.26 A plan obtained by the original beam arrangement (a) and five plans determined by similar-case-based beam arrangements (b)-(f) employed for quantifying the robustness of the beam directions. The relationship between the beam direction and the 0th moment of the power spectrum was derived for estimation of the robustness of each beam direction. We applied the proposed evaluation method to seven patients with head and neck cancer. The mean of the 0th moment in the conventional beam directions, which were empirically selected by a manual method, was statistically smaller than that for the avoided beam directions (p < 0.05), which means that the conventional beam directions based on planners’ experiences and knowledge were appropriate from the theoretical point of view. The results of this preliminary study may lead to an automated selection of beam directions based on the relationships mentioned above.