On-line re-optimization of prostate IMRT plans for adaptive radiation therapy.

For intermediate and high risk prostate cancer, both the prostate gland and seminal vesicles are included in the clinical target volume. Internal motion patterns of these two organs vary, presenting a challenge for adaptive treatment. Adaptive techniques such as isocenter repositioning and soft tissue alignment are effective when tumor volumes only exhibit translational shift, while direct re-optimization of the intensity-modulated radiation therapy (IMRT) plan maybe more desirable when extreme deformation or differential positioning changes of the organs occur. Currently, direct re-optimization of the IMRT plan using beamlet (or fluence map) has not been reported. In this study, we report a novel on-line re-optimization technique that can accomplish plan adjustment on-line. Deformable image registration is used to provide position variation information on each voxel along the three dimensions. The original planned dose distribution is used as the 'goal' dose distribution for adaptation and to ensure planning quality. Fluence maps are re-optimized via linear programming, and a plan solution can be achieved within 2 min. The feasibility of this technique is demonstrated with a clinical case with large deformation. Such on-line ART process can be highly valuable with hypo-fractionated prostate IMRT treatment.

Real-time inverse planning for Gamma Knife radiosurgery.

The challenges of real-time Gamma Knife inverse planning are the large number of variables involved and the unknown search space a priori. With limited collimator sizes, shots have to be heavily overlapped to form a smooth prescription isodose line that conforms to the irregular target shape. Such overlaps greatly influence the total number of shots per plan, making pre-determination of the total number of shots impractical. However, this total number of shots usually defines the search space, a pre-requisite for most of the optimization methods. Since each shot only covers part of the target, a collection of shots in different locations and various collimator sizes selected makes up the global dose distribution that conforms to the target. Hence, planning or placing these shots is a combinatorial optimization process that is computationally expensive by nature. We have previously developed a theory of shot placement and optimization based on skeletonization. The real-time inverse planning process, reported in this paper, is an expansion and the clinical implementation of this theory. The complete planning process consists of two steps. The first step is to determine an optimal number of shots including locations and sizes and to assign initial collimator size to each of the shots. The second step is to fine-tune the weights using a linear-programming technique. The objective function is to minimize the total dose to the target boundary (i.e., maximize the dose conformity). Results of an ellipsoid test target and ten clinical cases are presented. The clinical cases are also compared with physician's manual plans. The target coverage is more than 99% for manual plans and 97% for all the inverse plans. The RTOG PITV conformity indices for the manual plans are between 1.16 and 3.46, compared to 1.36 to 2.4 for the inverse plans. All the inverse plans are generated in less than 2 min, making real-time inverse planning a reality.

Center-crossing recurrent neural networks for the evolution of rhythmic behavior.

A center-crossing recurrent neural network is one in which the null-(hyper)surfaces of each neuron intersect at their exact centers of symmetry, ensuring that each neuron's activation function is centered over the range of net inputs that it receives. We demonstrate that relative to a random initial population, seeding the initial population of an evolutionary search with center-crossing networks significantly improves both the frequency and the speed with which high-fitness oscillatory circuits evolve on a simple walking task. The improvement is especially striking at low mutation variances. Our results suggest that seeding with center-crossing networks may often be beneficial, since a wider range of dynamics is more likely to be easily accessible from a population of center-crossing networks than from a population of random networks.