Fast VMAT planning for prostate radiotherapy: dosimetric validation of a deep learning-based initial segment generation method.

Objective. To develop and evaluate a deep learning based fast volumetric modulated arc therapy (VMAT) plan generation method for prostate radiotherapy.Approach. A customized 3D U-Net was trained and validated to predict initial segments at 90 evenly distributed control points of an arc, linked to our research treatment planning system (TPS) for segment shape optimization (SSO) and segment weight optimization (SWO). For 27 test patients, the VMAT plans generated based on the deep learning prediction (VMATDL) were compared with VMAT plans generated with a previously validated automated treatment planning method (VMATref). For all test cases, the deep learning prediction accuracy, plan dosimetric quality, and the planning efficiency were quantified and analyzed.Main results. For all 27 test cases, the resulting plans were clinically acceptable. TheV95%for the PTV2 was greater than 99%, and theV107%was below 0.2%. Statistically significant difference in target coverage was not observed between the VMATrefand VMATDLplans (P = 0.3243 > 0.05). The dose sparing effect to the OARs between the two groups of plans was similar. Small differences were only observed for the Dmean of rectum and anus. Compared to the VMATref, the VMATDLreduced 29.3% of the optimization time on average.Significance. A fully automated VMAT plan generation method may result in significant improvement in prostate treatment planning efficiency. Due to the clinically acceptable dosimetric quality and high efficiency, it could potentially be used for clinical planning application and real-time adaptive therapy application after further validation.

Anatomy structure creation and editing using 3D implicit surfaces.

PURPOSE: To accurately reconstruct, and interactively reshape 3D anatomy structures' surfaces using small numbers of 2D contours drawn in the most visually informative views of 3D imagery. The innovation of this program is that the number of 2D contours can be very much smaller than the number of transverse sections, even for anatomy structures spanning many sections. This program can edit 3D structures from prior segmentations, including those from autosegmentation programs. The reconstruction and surface editing works with any image modality. METHODS: Structures are represented by variational implicit surfaces defined by weighted sums of radial basis functions (RBFs). Such surfaces are smooth, continuous, and closed and can be reconstructed with RBFs optimally located to efficiently capture shape in any combination of transverse (T), sagittal (S), and coronal (C) views. The accuracy of implicit surface reconstructions was measured by comparisons with the corresponding expert-contoured surfaces in 103 prostate cancer radiotherapy plans. Editing a pre-existing surface is done by overdrawing its profiles in image views spanning the affected part of the structure, deleting an appropriate set of prior RBFs, and merging the remainder with the new edit contour RBFs. Two methods were devised to identify RBFs to be deleted based only on the geometry of the initial surface and the locations of the new RBFs. RESULTS: Expert-contoured surfaces were compared with implicit surfaces reconstructed from them over varying numbers and combinations of T/S/C planes. Studies revealed that surface-surface agreement increases monotonically with increasing RBF-sample density, and that the rate of increase declines over the same range. These trends were observed for all surface agreement metrics and for all the organs studied-prostate, bladder, and rectum. In addition, S and C contours may convey more shape information than T views for CT studies in which the axial slice thickness is greater than the pixel size. Surface editing accuracy likewise improves with larger sampling densities, and the rate of improvement similarly declines over the same conditions. CONCLUSIONS: Implicit surfaces based on RBFs are accurate representations of anatomic structures and can be interactively generated or modified to correct segmentation errors. The number of input contours is typically smaller than the number of T contours spanned by the structure.

Clinical validation of atlas-based auto-segmentation of multiple target volumes and normal tissue (swallowing/mastication) structures in the head and neck.

PURPOSE: To validate and clinically evaluate autocontouring using atlas-based autosegmentation (ABAS) of computed tomography images. METHODS AND MATERIALS: The data from 10 head-and-neck patients were selected as input for ABAS, and neck levels I-V and 20 organs at risk were manually contoured according to published guidelines. The total contouring times were recorded. Two different ABAS strategies, multiple and single subject, were evaluated, and the similarity of the autocontours with the atlas contours was assessed using Dice coefficients and the mean distances, using the leave-one-out method. For 12 clinically treated patients, 5 experienced observers edited the autosegmented contours. The editing times were recorded. The Dice coefficients and mean distances were calculated among the clinically used contours, autocontours, and edited autocontours. Finally, an expert panel scored all autocontours and the edited autocontours regarding their adequacy relative to the published atlas. RESULTS: The time to autosegment all the structures using ABAS was 7 min/patient. No significant differences were observed in the autosegmentation accuracy for stage N0 and N+ patients. The multisubject atlas performed best, with a Dice coefficient and mean distance of 0.74 and 2 mm, 0.67 and 3 mm, 0.71 and 2 mm, 0.50 and 2 mm, and 0.78 and 2 mm for the salivary glands, neck levels, chewing muscles, swallowing muscles, and spinal cord-brainstem, respectively. The mean Dice coefficient and mean distance of the autocontours vs. the clinical contours was 0.8 and 2.4 mm for the neck levels and salivary glands, respectively. For the autocontours vs. the edited autocontours, the mean Dice coefficient and mean distance was 0.9 and 1.6 mm, respectively. The expert panel scored 100% of the autocontours as a "minor deviation, editable" or better. The expert panel scored 88% of the edited contours as good compared with 83% of the clinical contours. The total editing time was 66 min. CONCLUSION: Multiple-subject ABAS of computed tomography images proved to be a useful novel tool in the rapid delineation of target and normal tissues. Although editing of the autocontours is inevitable, a substantial time reduction was achieved using editing, instead of manual contouring (180 vs. 66 min).

Atlas-based auto-segmentation of head and neck CT images.

Treatment planning for high precision radiotherapy of head and neck (H&N) cancer patients requires accurate delineation of many structures and lymph node regions. Manual contouring is tedious and suffers from large inter- and intra-rater variability. To reduce manual labor, we have developed a fully automated, atlas-based method for H&N CT image segmentation that employs a novel hierarchical atlas registration approach. This registration strategy makes use of object shape information in the atlas to help improve the registration efficiency and robustness while still being able to account for large inter-subject shape differences. Validation results showed that our method provides accurate segmentation for many structures despite difficulties presented by real clinical data. Comparison of two different atlas selection strategies is also reported.

Region segmentation using information divergence measures.

Image segmentations based on maximum likelihood or maximum a posteriori analyses of object textures usually assume parametric models (e.g., Gaussian) for distributions of these features. For real images, parameter accuracy and model stationarity may be elusive, so that model-free inference methods ought to have an advantage over those that are model-dependent. Functions of the relative entropy (RE) from information theory can produce minimum error, model-free inferences, and can detect the boundary of an image object by maximizing the RE between the pixel distributions inside and outside a flexible curve contour. A generalization of the RE -- the Jensen-Rényi divergence (JRD) -- computes optimal n-way decisions and can contour multiple objects in an image simultaneously. Seed regions expand naturally and multiple contours tend not to overlap. An edge detector based on the JRD, combined with multivariate pixel segmentation, generally improved the error of the segmentation. We apply these functions to contour patient anatomy in X-ray computed tomography for radiotherapy treatment planning.