Search

PURPOSE The purpose of this study is to evaluate the prediction and classification performances of the gamma passing rate (GPR) for different machine learning models and to select the best model for achieving machine learning-based patient-specific quality assurance (PSQA). METHODS The measurement verification of 356 head-and-neck volumetric modulated arc therapy plans was performed using a diode array phantom (Delta4 Phantom), and GPR values at 2%/2 mm with global normalization and 3%/2 mm with local normalization were calculated. Machine learning models, including ridge regression (RIDGE), random forest (RF), support vector regression (SVR), and stacked generalization (STACKING), were used to predict the GPR. Each machine learning model was trained using 260 plans, and the prediction accuracy was evaluated using the remaining 96 plans. The prediction error between the measured and predicted GPR was evaluated. For the classification evaluation, the lower control limit for the measured GPR and lower control limit for predicted GPR (LCLp ) was defined to identify whether the GPR values represent a "pass" or a "fail." LCLp values with 99% and 99.9% confidence levels were calculated as the upper prediction limits for the GPR estimated from the linear regression between the measured and predicted GPR. RESULTS There was an overestimation trend of the low measured GPR. The maximum prediction errors for RIDGE, RF, SVR, and STACKING were 3.2%, 2.9%, 2.3%, and 2.2% at global 2%/2 mm and 6.3%, 6.6%, 6.1%, and 5.5% at the local 3%/2 mm, respectively. In the global 2%/2 mm, the sensitivity was 100% for all the machine learning models except RIDGE when using 99% LCLp . The specificity was 76.1% for RIDGE, RF, and SVR and 66.3% for STACKING, however, the specificity decreased dramatically when 99.9% LCLp was used. In the local 3%/2 mm, however, only STACKING showed 100% sensitivity when using 99% LCLp . The decrease in the specificity using 99.9% LCLp was smaller than that in the global 2%/2mm, and the specificity for RIDGE, RF, SVR, and STACKING was 61.3%, 61.3%, 72.0%, and 66.8%, respectively. CONCLUSIONS STACKING had better prediction accuracy for low GPR values than other machine learning models. Applying LCLp to a regression model enabled the consistent evaluation of quantitative and qualitative GPR predictions. Adjusting the confidence level of the LCLp helped improve the balance between the sensitivity and specificity. We suggest that STACKING can assist the safe and efficient operation of PSQA. This article is protected by copyright. All rights reserved.

The calibration coefficients of a parallel plate ionization chamber are examined by comparing the coefficients obtained through three methods: a calculation from a 60Co calibration coefficient, N(D, omega, 60Co), a cross-calibration of a parallel plate ionization chamber using a cylindrical ionization chamber at the plateau region of a mono-energetic beam and a cross-calibration of the chamber using a cylindrical chamber at the middle of the SOBP of the therapeutic beams. This paper also examines reference conditions for determining absorbed dose to water in the cases of therapeutic carbon and proton beams. In the dose calibration procedure recommended by IAEA, irradiation fields should be larger than 10 cm in diameter and the water phantom should extend by at least 5 cm beyond each side of the field. These recommendations are experimentally verified for proton and carbon beams. For proton beams, the calibration coefficients obtained by these three methods approximately agreed. For carbon beams, the calibration coefficients obtained by the second method were about 1.0% larger than those obtained by the third method, and the calibration coefficients obtained by cross-calibration using 290 MeV/u beams were 0.5% lower than those obtained using 400 MeV/u beams. The calibration coefficient obtained by the first method agreed roughly with the results obtained by SOBP beams.

The purpose of this study was to obtain a set of correction factors of the radiophotoluminescent glass dosimeter (RGD) output for field size changes and wedge insertions.Several linear accelerators were used for irradiation of the RGDs. The field sizes were changed from 5 × 5 cm to 25 × 25 cm for 4, 6, 10, and 15 MV x-ray beams. The wedge angles were 15°, 30°, 45°, and 60°. In addition to physical wedge irradiation, nonphysical (dynamic/virtual) wedge irradiations were performed.The obtained data were fitted with a single line for each energy, and correction factors were determined. Compared with ionization chamber outputs, the RGD outputs gradually increased with increasing field size, because of the higher RGD response to scattered low-energy photons. The output increase was about 1% per 10 cm increase in field size, with a slight difference dependent on the beam energy. For both physical and nonphysical wedged beam irradiation, there were no systematic trends in the RGD outputs, such as monotonic increase or decrease depending on the wedge angle change if the authors consider the uncertainty, which is approximately 0.6% for each set of measured points. Therefore, no correction factor was needed for all inserted wedges. Based on this work, postal dose audits using RGDs for the nonreference condition were initiated in 2010. The postal dose audit results between 2010 and 2012 were analyzed. The mean difference between the measured and stated doses was within 0.5% for all fields with field sizes between 5 × 5 cm and 25 × 25 cm and with wedge angles from 15° to 60°. The standard deviations (SDs) of the difference distribution were within the estimated uncertainty (1SD) except for the 25 × 25 cm field size data, which were not reliable because of poor statistics (n = 16).A set of RGD output correction factors was determined for field size changes and wedge insertions. The results obtained from recent postal dose audits were analyzed, and the mean differences between the measured and stated doses were within 0.5% for every field size and wedge angle. The SDs of the distribution were within the estimated uncertainty, except for one condition that was not reliable because of poor statistics.

We evaluated the CT number accuracy in determining a CT calibration method for treatment planning by use of a 256 multi-slice CT (MSCT) scanner. An electron density phantom, which extends full length in the longitudinal direction, was scanned by the 256 MSCT scanner in a single rotation. We inserted four types of samples (air, 100 % ethanol, 40 wt% aqueous K(2)HPO(4), and water) into the phantom. The regions of interest were set for all image slices, and the average CT number was calculated in the transverse and longitudinal sections. For the transverse image direction, the CT number became lower toward the center of the phantom with almost all samples. The causes of these phenomena might be attributed to the effects of scattered radiation and those of beam hardening due to the high-atomic-number material (40 wt% aqueous K(2)HPO(4)). However, the phenomena were too complicated to allow us to determine their causes only from the present studies. Meanwhile, for increasing the accuracy of the CT number calibration, a single sample should be inserted, and this step should be repeated with different samples. For the longitudinal direction, the CT number for a 40 wt % aqueous K(2)HPO(4) sample increased by 30 HU from the anode to the cathode side due to variations in the X-ray energy caused by the heel effect. The caveat is that the CT number varies in the longitudinal direction when a CT scanner with a wide beam width is used.

The postal dose audit using radio-photoluminescence glass dosimeters was begun in November 2007 in order to improve the quality of radiotherapy in Japan. However, the irradiation conditions are now limited to the reference conditions which are 10×10 cm(2) field and 10 cm depth. The application of the glass dosimeters to non-reference conditions is strongly desired. This study dealt with the field-size dependence of the glass dosimeter outputs in the 6 MV photon beams of a medical linear accelerator (Varian Clinac21EX). We irradiated glass dosimeters with square field sizes of 5, 7, 10, 13, 16, 20, 23 and 25 cm side lengths at the 10 cm depth of the water equivalent phantom (SSD=90 cm). The outputs were compared with ionization chamber outputs. The ratio of the glass dosimeter outputs to the absorbed dose to water obtained with the ionization chamber increased approximately 1.5% between 5×5 cm(2) and 25×25 cm(2). We have to consider this field-size dependence when we apply the glass dosimeters to non-reference conditions.

To estimate the physical dose at the center of spread-out Bragg peaks (SOBP) for various conditions of the irradiation system, a semiempirical approach was applied. The dose at the center of the SOBP depends on the field size because of large-angle scattering particles in the water phantom. For a small field of 5 x 5 cm2, the dose was reduced to 99.2%, 97.5%, and 96.5% of the dose used for the open field in the case of 290, 350, and 400 MeV/n carbon beams, respectively. Based on the three-Gaussian form of the lateral dose distributions of the carbon pencil beam, which has previously been shown to be effective for describing scattered carbon beams, we reconstructed the dose distributions of the SOBP beam. The reconstructed lateral dose distribution reproduced the measured lateral dose distributions very well. The field-size dependencies calculated using the reconstructed lateral dose distribution of the therapeutic carbon beam agreed with the measured dose dependency very well. The reconstructed beam was also used for irregularly shaped fields. The resultant dose distribution agreed with the measured dose distribution. The reconstructed beams were found to be applicable to the treatment-planning system.

In order to support a routine QA of the CT number for treatment planning, we developed a phantom and a sample holder for easy handling. At most particle radiotherapy facilities in Japan, the CT number is calibrated by the poly-binary calibration method using liquid samples of 100% ethanol and 40% K(2)HPO(4) which are set in a cylindrical water phantom. However it is hard to remove air bubbles from the calibration liquid sample and maintain its stable concentration for a long time. So much time is needed for QA of the CT number. The new sample holder, which we developed, was able to keep a stable concentration of the liquid for more than 300 days. Consequently, the CT number of each sample, which was set in a water equivalent solid phantom, was the same as the CT number in a water phantom within 7 HU. In addition, we developed software which could measure the CT number of each sample semi-automatically and could calculate the calibration coefficients between the CT number and water equivalent length (WEL). Using this software, we could check the calibration result instantly at the time of CT data acquisition. These tools should be useful to carry out calibration of the CT-WEL routinely in a short time.