No abstract

This paper proposes a new geometrical formulation of the coplanar beam orientation problem combined with a hybrid multiobjective genetic algorithm. The approach is demonstrated by optimizing the beam orientation in two dimensions, with the objectives being formulated using planar geometry. The traditional formulation of the objectives associated with the organs at risk has been modified to account for the use of complex dose delivery techniques such as beam intensity modulation. The new algorithm attempts to replicate the approach of a treatment planner whilst reducing the amount of computation required. Hybrid genetic search operators have been developed to improve the performance of the genetic algorithm by exploiting problem-specific features. The multiobjective genetic algorithm is formulated around the concept of Pareto optimality which enables the algorithm to search in parallel for different objectives. When the approach is applied without constraining the number of beams, the solution produces an indication of the minimum number of beams required. It is also possible to obtain non-dominated solutions for various numbers of beams, thereby giving the clinicians a choice in terms of the number of beams as well as in the orientation of these beams.

Image-guided radiation therapy aims to improve the accuracy of treatment delivery by tracking tumor position and compensating for observed movement. Due to system latency it is sometimes necessary to predict tumor trajectory evolution in order to facilitate changes in beam delivery. Neural networks (NNs) have previously been investigated for predicting future tumor position because of their ability to model non-linear systems. However, no attempt has been made to optimize the NN training algorithms, and no mention has been made of potential errors which can be caused by using NNs for extrapolation purposes. In this work, after giving a brief explanation of NN theory, a comparison is made between 4 different adaptive algorithms for training time-series prediction NNs. New error criteria are introduced which highlight error maxima. Results are obtained by training the NNs using previously published data. A hybrid algorithm combining Bayesian regularization with conjugate-gradient backpropagation is demonstrated to give the best average prediction accuracy, whilst a generalized regression NN is shown to reduce the possibility of isolated large prediction errors.

Respiration induces significant movement of tumours in the vicinity of thoracic and abdominal structures. Real-time image-guided radiotherapy (IGRT) aims to adapt radiation delivery to tumour motion during irradiation. One of the main problems for achieving this objective is the presence of time lag between the acquisition of tumour position and the radiation delivery. Such time lag causes significant beam positioning errors and affects the dose coverage. A method to solve this problem is to employ an algorithm that is able to predict future tumour positions from available tumour position measurements. This paper presents a multiple model approach to respiratory-induced tumour motion prediction using the interacting multiple model (IMM) filter. A combination of two models, constant velocity (CV) and constant acceleration (CA), is used to capture respiratory-induced tumour motion. A Kalman filter is designed for each of the local models and the IMM filter is applied to combine the predictions of these Kalman filters for obtaining the predicted tumour position. The IMM filter, likewise the Kalman filter, is a recursive algorithm that is suitable for real-time applications. In addition, this paper proposes a confidence interval (CI) criterion to evaluate the performance of tumour motion prediction algorithms for IGRT. The proposed CI criterion provides a relevant measure for the prediction performance in terms of clinical applications and can be used to specify the margin to accommodate prediction errors. The prediction performance of the IMM filter has been evaluated using 110 traces of 4-minute free-breathing motion collected from 24 lung-cancer patients. The simulation study was carried out for prediction time 0.1-0.6 s with sampling rates 3, 5 and 10 Hz. It was found that the prediction of the IMM filter was consistently better than the prediction of the Kalman filter with the CV or CA model. There was no significant difference of prediction errors for the sampling rates 5 and 10 Hz. For these sampling rates, the errors of the IMM filter for 0.4 s prediction time were less than 2.1 mm in terms of the 95% CI criterion or 1.1 mm in terms of the standard deviation (SD) or root mean squared errors (RMSE) criterion. For the prediction time of 0.6 s the errors were less than 3.6 mm in terms of the 95% CI criterion or 1.8 mm in terms of the SD/RMSE criterion. The prediction error analysis showed that the average percentage of the target lies outside the 95% CI margin was 5.2% and outside the SD/RMSE margin was 24.3%. This indicates the effectiveness of the 95% CI criterion as a margining strategy to accommodate prediction errors.

In this paper a new method is proposed to quantify and reduce the radiation beam position uncertainty due to the radiotherapy treatment machine gantry deflection. A new tool has been designed and manufactured to provide the means to measure the alignment of the collimator axis and of the beam central axis in space, using the NDI Polaris optical tracking system and Gafchromic films. The tool can be mounted onto the accessory tray of the linacs from different manufacturers. The approach has been demonstrated with measurements of the mechanical isocentre being performed on ten linacs from three major manufacturers at four clinical sites. Measurements of the radiation isocentre were performed on a single linac. The collimator axis trajectory is modelled using a vector-end effector in order to provide more information than standard mechanical assessment methods. The method uses a mathematical optimization technique to calculate the position of the mechanical isocentre and the 'size' of the collimator axes intersection locus. Deviations of the collimator axes from the isocentre are expressed in terms of systematic and random error. The effects of measurement uncertainties are evaluated both via simulations and experimentally. The use of optical tracking and optimization techniques combined with an operator-induced measurement error compensation algorithm leads to a faster measurement of the mechanical isocentre (20 min for 24 angles) and eliminates operator-induced uncertainties. The uncertainty of the measurement of the mechanical isocentre was between 40 microm and 70 microm in terms of standard deviation. For some of the linacs assessed, the mechanical isocentre obtained using a standard approach with an adjustable pointer was displaced by over 1 mm from that found with the proposed method. The radii of the collimator axes intersection locus found with the proposed method were between 0.4 mm and 0.72 mm for the linacs assessed. Film measurement revealed a misalignment of the mechanical isocentre and the radiation isocentre by 0.4 mm. The proposed method potentially enables a reduction of the beam position uncertainty. This may be achieved at the planning stage by compensating for the identified systematic collimator axes deviations which were found to be reproducible. The method also creates a potential for using different setup margins independently for each axis and for each gantry angle, calculated specifically for a given linac.