Predicting the parameters of a prostate IMRT objective function based on dose statistics under fixed parameter settings Renzhi Lu, Richard J. Radke 1, Andrew Jackson 2 Rensselaer Polytechnic Institute 1,Memorial Sloan-Kettering Cancer Center 2 Abstract To simplify the trial-and-error process of adjusting objective function parameters (e.g. weights, dose limits) in prostate IMRT planning, we present a feasibility study showing that machine learning followed by a sensitivity-driven greedy search can quickly and automatically determine parameters that lead to a plan meeting the clinical requirements. The training database is composed of a large number of plans treated effectively under the 8640cGy prostate IMRT protocol. For each plan, the output features include the clinical setting of parameters, and the input features include simple dose statistics resulting from several fixed settings of parameters. We predict a new plan based on the 3 nearest plans in the training database that have similar input features. Starting from such a pre-plan, a sensitivity based automatic parameter search is applied to improve the plan’s deficiencies. Experiments on a 39-patient dataset showed that a clinically acceptable (based on simplified dose calculation) prostate IMRT plan could be automatically determined within 2 minutes of optimization. State-of-the-art Hunt et al. gave specified procedure for changes to be made in optimization parameters given specific deficits in plans. Barbiere et al. looked for the best optimization parameters via structured grid searches over historical plans. Berger et al. utilized the parameters of similar cases in the historical database for a new patient. The similarity measure was based on an ellipsoidal patient geometry. Challenges and significance Planning a prostate IMRT case can take many hours, even for an expert planner. The bottleneck of current IMRT systems is not optimization, but the trial and error procedure of adjusting optimization parameters. Circumventing or minimizing this procedure would save many person-hours of effort. Technical approach 1. Problem description Fig. 1 Prostate IMRT: visualization of beams and structures Input : Settings for 5 beams. Contours for 6 structures. Optimization : : find the best radiation intensities I for a fixed parameter set P. Adjustment : Change the parameters P, redo optimization. Clinical goal: Dose D in target is as prescribed, in normal tissues is minimized. Our task: Find a set of P* such that minimizing the corresponding objective function meets all clinical goals. Fig.3 C omparison of dose evaluation statistics (based on simplified dose calculation) for: Upper: predicted pre-plans versus clinical plans. Lower: adjusted plans versus clinical plans Left to right: PTV V95, rectwall V54, rectwall Dmax Fig. 4 Comparison of dose volume histogram (based on simplified dose calculation) for predicted plans after adjustment vs clinical plans. Left: Patient 1. Right: Patient 2. We conclude that machine learning that utilizes the knowledge of past plans can predict a good starting point for parameter selection in IMRT. Given the deficiencies of the prediction, a sensitivity-driven greedy algorithm can effectively automate the necessary adjustment. Publications Ack. NSF Support :R. Lu et al, “Reduced-order parameter optimization for prostate Intensity- modulated Radiotherapy”, Phy in Med & Bio, Vol. 52, , Feb 2007 R. Lu, Richard Radke et al, “Learning the Relationship between Patient Geometry and Beam Intensity in Breast Intensity-Modulated Radiotherapy,” IEEE Trans. on Biom Eng, Vol. 53, No. 5, pp , May Future Plans Use machine learning to predict the intensities for general IMRT planning References 1. Hunt et al, “Evaluation of concave dose distributions created using an inverse planning system,” Int J Radiat Oncol Biol Phys, Vol 54, pp , J. Barbiere et al, “ A parameter optimization algorithm for intensity- modulated radiotherapy prostate treatment planning.”, J of Applied Clinical Med Phy, 3:227–234, J. Berger, Roentgen: radiation therapy and case-based reasoning Proceedings of the Tenth Conference on Artificial Intelligence, ,1994. Contact information Richard J. Radke, Assistant Professor Dept. of Electrical, Computer, and Systems Engineering Rensselaer Polytechnic Institute th Street, Troy, NY phone: (518) , This work was supported in part by Gordon-CenSSIS, the Bernard M. Gordon Center for Subsurface Sensing and Imaging Systems, under the Engineering Research Centers Program of the National Science Foundation (Award Number EEC ). 2. System overview Fig. 2 Overview of our two-step approach 3. Generating a pre-plan using machine learning Given a large number of historical plans, we use K-Nearest-Neighbor method that locates the K(=3) “closest” training cases and predicts P* based on a weighted average of the K training outputs. Input features: dose statistics D i under 3 fixed parameter settings of PTV{V95, Dmin}, rectum wall {V54,V87, Dmax}, and bladder wall V54. Output features: optimization parameters P Feature weighting: The input statistics may have different influences for each output parameter. We weight each statistic differently according to the correlation coefficients between Di and P in the training data, shown in Table 1. KNN regression is based on the K nearest training samples with output Pi and Euclidean distance di: Table 1. PEAR correlations between output parameters in P* (columns) and input statistics at default parameters P0 (rows) 4. Sensitivity driven greedy search Given the deficiencies in the pre-plan, determine which parameters should be tuned (by global sensitivity) and to which extent (by local sensitivity). Global sensitivity: Pearson correlation coefficient for the input/output pairs from Monte Carlo simulation Local sensitivity: Forward differentiation Greedy search: Each iteration updates the most sensitive parameter Pi to be: 5. Results In Figure 3(upper row), the PTV coverage for the pre-plans generally agrees with the ground truth and satisfies the clinical criteria. However, several predicted plans violate the rectum constraint boundaries. In Figure 3 (lower row), after greedy search the adjusted plan greatly improves the dose deficiencies in the rectum. The mean absolute differences across 39 patients for PTV V95, rectwall V54 and Dmax are 1.1%, 2.6% and 0.9%, respectively. Figure 4 compares sample DVHs for two cases. The adjusted plans both satisfy the clinical constraints.