Purposeful irradiation of the epidural space to enhance local control without compromising cord sparing in spine radiosurgery†.

Purpose: The epidural space is a frequent site of cancer recurrence after spine stereotactic radiosurgery (SSRS). This may be due to microscopic disease in the epidural space which is underdosed to obey strict spinal cord dose constraints. We hypothesized that the epidural space could be purposefully irradiated to prescription dose levels, potentially reducing the risk of recurrence in the epidural space without increasing toxicity. Methods and materials: SSRS clinical treatment plans with spinal cord contours, spinal planning target volumes (PTVspine), and delivered dose distributions were retrospectively identified. An epidural space PTV (PTVepidural) was contoured to avoid the spinal cord and focus on regions near the PTVspine. Clinical plan constraints included PTVspine constraints (D95% and D5%, based on prescription dose) and spinal cord constraints (Dmax < 1300 cGy, D10% < 1000 cGy). Plans were revised with three prescriptions of 1800, 2000 and 2400 cGy in two sets, with one set of revisions (supplemented plans) designed to additionally target the PTVepidural by optimizing PTVepidural D95% in addition to meeting every clinical plan constraint. Clinical and revised plans were compared according to their PTVepidural DVH distributions, and D95% distributions. Results: Seventeen SSRS plans meeting the above criteria were identified. Supplemented plans had higher doses to the epidural low-dose regions at all prescription levels. Epidural PTV D95% values for the supplemented plans were all statistically significantly different from the values of the base plans (p < 10-4). The epidural PTV D95% increases depended on the initial prescription, increasing from 11.52 to 16.90 Gy, 12.23 to 18.85 Gy, and 13.87 to 19.54 Gy for target prescriptions of 1800, 2000 and 2400 cGy, respectively. Conclusions: Purposefully targeting the epidural space in SSRS may increase control in the epidural space without significantly increasing the risk of spinal cord toxicity. A clinical trial of this approach should be considered.

A Novel Machine Learning Model for Dose Prediction in Prostate Volumetric Modulated Arc Therapy Using Output Initialization and Optimization Priorities.

Treatment planning for prostate volumetric modulated arc therapy (VMAT) can take 5-30 min per plan to optimize and calculate, limiting the number of plan options that can be explored before the final plan decision. Inspired by the speed and accuracy of modern machine learning models, such as residual networks, we hypothesized that it was possible to use a machine learning model to bypass the time-intensive dose optimization and dose calculation steps, arriving directly at an estimate of the resulting dose distribution for use in multi-criteria optimization (MCO). In this study, we present a novel machine learning model for predicting the dose distribution for a given patient with a given set of optimization priorities. Our model innovates upon the existing machine learning techniques by utilizing optimization priorities and our understanding of dose map shapes to initialize the dose distribution before dose refinement via a voxel-wise residual network. Each block of the residual network individually updates the initialized dose map before passing to the next block. Our model also utilizes contiguous and atrous patch sampling to effectively increase the receptive fields of each layer in the residual network, decreasing its number of layers, increasing model prediction and training speed, and discouraging overfitting without compromising on the accuracy. For analysis, 100 prostate VMAT cases were used to train and test the model. The model was evaluated by the training and testing errors produced by 50 iterations of 10-fold cross-validation, with 100 cases randomly shuffled into the subsets at each iteration. The error of the model is modest for this data, with average dose map root-mean-square errors (RMSEs) of 2.38 ± 0.47% of prescription dose overall patients and all optimization priority combinations in the patient testing sets. The model was also evaluated at iteratively smaller training set sizes, suggesting that the model requires between 60 and 90 patients for optimal performance. This model may be used for quickly estimating the Pareto set of feasible dose objectives, which may directly accelerate the treatment planning process and indirectly improve final plan quality by allowing more time for plan refinement.

An Interpretable Planning Bot for Pancreas Stereotactic Body Radiation Therapy.

PURPOSE: Pancreas stereotactic body radiation therapy (SBRT) treatment planning requires planners to make sequential, time-consuming interactions with the treatment planning system to reach the optimal dose distribution. We sought to develop a reinforcement learning (RL)-based planning bot to systematically address complex tradeoffs and achieve high plan quality consistently and efficiently. METHODS AND MATERIALS: The focus of pancreas SBRT planning is finding a balance between organ-at-risk sparing and planning target volume (PTV) coverage. Planners evaluate dose distributions and make planning adjustments to optimize PTV coverage while adhering to organ-at-risk dose constraints. We formulated such interactions between the planner and treatment planning system into a finite-horizon RL model. First, planning status features were evaluated based on human planners' experience and defined as planning states. Second, planning actions were defined to represent steps that planners would commonly implement to address different planning needs. Finally, we derived a reward system based on an objective function guided by physician-assigned constraints. The planning bot trained itself with 48 plans augmented from 16 previously treated patients, and generated plans for 24 cases in a separate validation set. RESULTS: All 24 bot-generated plans achieved similar PTV coverages compared with clinical plans while satisfying all clinical planning constraints. Moreover, the knowledge learned by the bot could be visualized and interpreted as consistent with human planning knowledge, and the knowledge maps learned in separate training sessions were consistent, indicating reproducibility of the learning process. CONCLUSIONS: We developed a planning bot that generates high-quality treatment plans for pancreas SBRT. We demonstrated that the training phase of the bot is tractable and reproducible, and the knowledge acquired is interpretable. As a result, the RL planning bot can potentially be incorporated into the clinical workflow and reduce planning inefficiencies.

Technical note: Interpolated Pareto surface similarity metrics for multi-criteria optimization in radiation therapy.

PURPOSE: There is a strong clinical need to evaluate different multi-criteria optimization (MCO) algorithms, including inverse optimization sampling algorithms and machine learning-based predictions. This study aims to develop and compare several interpolated Pareto surface similarity metrics. MATERIALS AND METHODS: The first metric is the root-mean-square error (RMSE) evaluated between vertices on the interpolated surfaces, augmented by intra-simplex sampling of the barycentric coordinates of the surfaces' simplicial complexes. The second metric is the average projected distance (APD), which evaluates the displacements between the vertices and computes their projections along the mean displacement. The third metric is the average nearest-point distance (ANPD), which numerically integrates point-to-simplex distances over the sampled simplices of the interpolated surfaces. These metrics were compared by their convergence rates, the times required to achieve convergence, and their representation of the underlying surface interpolations. For analysis, several interpolated Pareto surface pairs were constructed abstractly, with one pair from a nasopharyngeal treatment planning case using MCO. RESULTS: Convergence within 1% is typically achieved at approximately 50 and 80 samples per barycentric dimension for the RMSE and the ANPD, respectively. Calculation requires approximately 1 and 10 ms to achieve convergence for the RMSE and the ANPD in two dimensions, respectively, while the APD always requires < 1 ms. These time costs are much higher in higher dimensions for just the RMSE and ANPD. The APD values more closely approximated the ANPD limits than the RMSE limits. CONCLUSION: The ANPD's formulation and generality make it likely more meaningful than the RMSE and APD for representing the similarity between the underlying interpolated surfaces rather than the sampling points on the surfaces. However, in situations requiring high-speed evaluations, the APD may be more desirable due to its speed, independence from a subjectively chosen sampling rate, and similarity to the ANPD limits.