| Autor: |
Lau, Kenneth, Nordström, Håkan, Silva, Joakim da, Hennix, Marcus, Erlandsson, Joakim Wang, Körting, Clara, Dahlberg, John, Somell, Björn |
| Předmět: |
|
| Zdroj: |
Journal of Radiosurgery & SBRT; 2022 Supplement, Vol. 8, p161-162, 2p |
| Abstrakt: |
Isocenter placement is important in creating Gamma Knife treatment plans. The recently released inverse planner – Leksell Gamma Knife® (LGK) Lightning (Elekta AB, Sweden) has an isocenter placement algorithm that automatically generates a set of isocenters based on the target’s geometric description. The algorithm has generally shown very good performance. However, to limit complexity, the number of isocenters is limited, resulting in suboptimal plans for some large targets. Exploring the entire space of possible isocenters would likely allow for better plans but would require solving an enormously expensive optimization problem, making it undesirable in clinical workflow. We propose a datadriven approach that generates a set of isocenter locations without exploring the entire search space at inference. Instead of running a larger optimization problem, we use a machine learning (ML) model to predict the best possible isocenters based on training data. We first increase the search space by relaxing the constraints on the number of isocenter candidates (degrees of freedom) in the optimization problem. The corresponding (timeconsuming) LGK Lightning optimization is then run offline for many cases, identifying and storing isocenter positions surviving the optimization for each case. An ML model is then trained to map the surviving isocenter sets from target feature descriptors. We use a convolutional neural network that resembles U-Net. We generated isocenters from a dataset of 2921 cases with a large variation in size and shape, and split them into a training set (2337 cases) and a validation set (584 cases). We evaluated LGK Lightning-optimized plans from both the relaxed isocenter placement (max-out) and the ML algorithm against those from the current isocenter placement algorithm. We compared their results on Paddick conformity index (PCI) and beam-ontime (BOT) with different low-dose and BOT weights; and the number of proposed isocenters for the optimization problem. Our results show that the max-out algorithm results in better dose conformity but longer BOT than the current fill algorithm in LGK Lightning. For tumors larger than 3cm3, the PCI is 2.13% (+-1.74%) better with 5.23% (+-14.35%) longer BOT but 904% (+-474%) more isocenters. The ML model shows similar behavior: the PCI is 0.64% (+-2.18%) better with 6.51% (+-13.4%) longer BOT and has 184% (+-125%) more isocenters than the current fill algorithm. [ABSTRACT FROM AUTHOR] |
| Databáze: |
Complementary Index |
| Externí odkaz: |
|
