Logo Logo
Help
Contact
Switch language to German
Towards EPID-based 3D in vivo dosimetry for modern radiation therapy
Towards EPID-based 3D in vivo dosimetry for modern radiation therapy
Modern radiotherapy techniques, such as Intensity Modulated Radiation Therapy (IMRT) and Volumetric Modulated Arc Therapy (VMAT), can deliver highly conformal dose distributions, with steep dose gradients between the target and organs at risk. This increases the demands on proper quality assurance and dose verification before (pre-treatment) and during (in vivo) patient irradiation. This project proposes a methodology for EPID-based in vivo dosimetry, combining the accuracy of Monte Carlo (MC) methods for dose simulation in patient geometry, with the time-efficiency of deep neural networks. The Deep Dose Estimation (DDE) network, originally developed for dose estimation in radiological computed tomography (CT) exams, has been extended and trained to predict 3D dose distributions due to IMRT fields, inside a patient, with accuracy comparable to MC methods. The DDE uses as input a patient CT image and an approximated dose distribution, called first order dose approximation (FOD), reconstructed from simulated EPID signals. The network was trained to map this two-channel input to an accurate dose distribution (ADD) inside the same patient CT, simulated using MC methods. The FODs are simplified 3D dose distributions produced as backprojections of the simulated EPID signals, accounting for magnification and inverse square law corrections, and attenuation through the virtual patient model. The FODs do not account for several effects, such as the build-up, beam hardening and scattering within the patient, all of which were properly considered in the ADDs. Hence, the methodology relies strongly on the MC model used to produce both the ADD and the transmitted EPID signals. A reliable MC model of the linac considered in this work was constructed and extensively validated. The patient-dependent part of the linac head, namely the multi-leaf collimator (MLC) system, was produced based entirely on information available in the literature. A virtual model of the EPID was also included in the patient-dependent part, to simultaneously record the transmitted signal through the virtual patient. The patient-independent part, i.e. the static parts of the linac head, was constructed based on confidential information provided by the vendor, and used to produce phase space (PhSp) files. These PhSp files were subsequently used as primary particle generators to simulate the ADDs and EPID signals. An alternative methodology for optimization of existing IAEA PhSp files was developed as a side project, which can be used to model the patient-independent part of the linac head when confidential vendor information is not available. The ADDs for clinical prostate IMRT fields, and respective transmitted EPID signals, were simulated inside 83 pelvic CTs, with gantry at 0�. In total, 581 different ADD-FOD sets were produced, with seven different fields per patient CT. The network was trained using the data sets of 67 patients (training set). The data of the remaining 16 patients were used for validation (test set). An additional dataset with eight fields simulated with gantry at 90� (lateral set) was used for evaluating the performance of the trained DDE for other irradiation directions. The quality of the DDE-predicted dose distributions (DDEP) on the test and lateral sets was quantified in terms of the gamma analysis with respect to the ADD (3%, 2 mm criteria). To evaluate the improvement obtained with the DDE, the same evaluation was performed for FODs and respective ADDs. The gamma passing rates between FODs and ADDs were as low as 46%, while for DDEPs the passing rates were above 97% for all fields on the test set. For the fields in the lateral set, the DDE was able to improve the passing rates from 88% to above 95%. The high passing rates for DDEPs indicate that the DDE was able to convert the FODs into ADDs, properly accounting for all missing effects. Moreover, once trained, the DDE can predict the dose inside a patient CT within 0.6 s per field (using a GPU), in contrast to 14 h needed for MC simulations (using a CPU-cluster). The dose delivered to a patient due to an entire prostate treatment session can therefore be predicted in less than one minute. With the proposed methodology, 3D in vivo dose distributions due to clinical patient irradiation can be obtained within seconds, potentially paving the way towards a clinically viable, real-time EPID-based in vivo dosimetry.
Not available
Martins, Juliana Cristina
2023
English
Universitätsbibliothek der Ludwig-Maximilians-Universität München
Martins, Juliana Cristina (2023): Towards EPID-based 3D in vivo dosimetry for modern radiation therapy. Dissertation, LMU München: Faculty of Physics
[thumbnail of Martins_Juliana_Cristina.pdf]
Preview
PDF
Martins_Juliana_Cristina.pdf

81MB

Abstract

Modern radiotherapy techniques, such as Intensity Modulated Radiation Therapy (IMRT) and Volumetric Modulated Arc Therapy (VMAT), can deliver highly conformal dose distributions, with steep dose gradients between the target and organs at risk. This increases the demands on proper quality assurance and dose verification before (pre-treatment) and during (in vivo) patient irradiation. This project proposes a methodology for EPID-based in vivo dosimetry, combining the accuracy of Monte Carlo (MC) methods for dose simulation in patient geometry, with the time-efficiency of deep neural networks. The Deep Dose Estimation (DDE) network, originally developed for dose estimation in radiological computed tomography (CT) exams, has been extended and trained to predict 3D dose distributions due to IMRT fields, inside a patient, with accuracy comparable to MC methods. The DDE uses as input a patient CT image and an approximated dose distribution, called first order dose approximation (FOD), reconstructed from simulated EPID signals. The network was trained to map this two-channel input to an accurate dose distribution (ADD) inside the same patient CT, simulated using MC methods. The FODs are simplified 3D dose distributions produced as backprojections of the simulated EPID signals, accounting for magnification and inverse square law corrections, and attenuation through the virtual patient model. The FODs do not account for several effects, such as the build-up, beam hardening and scattering within the patient, all of which were properly considered in the ADDs. Hence, the methodology relies strongly on the MC model used to produce both the ADD and the transmitted EPID signals. A reliable MC model of the linac considered in this work was constructed and extensively validated. The patient-dependent part of the linac head, namely the multi-leaf collimator (MLC) system, was produced based entirely on information available in the literature. A virtual model of the EPID was also included in the patient-dependent part, to simultaneously record the transmitted signal through the virtual patient. The patient-independent part, i.e. the static parts of the linac head, was constructed based on confidential information provided by the vendor, and used to produce phase space (PhSp) files. These PhSp files were subsequently used as primary particle generators to simulate the ADDs and EPID signals. An alternative methodology for optimization of existing IAEA PhSp files was developed as a side project, which can be used to model the patient-independent part of the linac head when confidential vendor information is not available. The ADDs for clinical prostate IMRT fields, and respective transmitted EPID signals, were simulated inside 83 pelvic CTs, with gantry at 0�. In total, 581 different ADD-FOD sets were produced, with seven different fields per patient CT. The network was trained using the data sets of 67 patients (training set). The data of the remaining 16 patients were used for validation (test set). An additional dataset with eight fields simulated with gantry at 90� (lateral set) was used for evaluating the performance of the trained DDE for other irradiation directions. The quality of the DDE-predicted dose distributions (DDEP) on the test and lateral sets was quantified in terms of the gamma analysis with respect to the ADD (3%, 2 mm criteria). To evaluate the improvement obtained with the DDE, the same evaluation was performed for FODs and respective ADDs. The gamma passing rates between FODs and ADDs were as low as 46%, while for DDEPs the passing rates were above 97% for all fields on the test set. For the fields in the lateral set, the DDE was able to improve the passing rates from 88% to above 95%. The high passing rates for DDEPs indicate that the DDE was able to convert the FODs into ADDs, properly accounting for all missing effects. Moreover, once trained, the DDE can predict the dose inside a patient CT within 0.6 s per field (using a GPU), in contrast to 14 h needed for MC simulations (using a CPU-cluster). The dose delivered to a patient due to an entire prostate treatment session can therefore be predicted in less than one minute. With the proposed methodology, 3D in vivo dose distributions due to clinical patient irradiation can be obtained within seconds, potentially paving the way towards a clinically viable, real-time EPID-based in vivo dosimetry.