Significant efforts such as the Pediatric Proton/Photon Consortium Registry (PPCR) involving multiple proton therapy centers have been made to conduct collaborative studies evaluating outcomes following proton therapy. As a groundwork dosimetry effort for the late effect investigation, we developed a Monte Carlo (MC) model of proton pencil beam scanning (PBS) to estimate organ/tissue doses of pediatric patients at the Maryland Proton Treatment Center (MPTC), one of the proton centers involved in the PPCR. The MC beam modeling was performed using the TOPAS (TOol for PArticle Simulation) MC code and commissioned to match measurement data within 1% for range, and 0.3 mm for spot sizes. The established MC model was then tested by calculating organ/tissue doses for sample intracranial and craniospinal irradiations on whole-body pediatric computational human phantoms. The simulated dose distributions were compared with the treatment planning system dose distributions, showing the 3 mm/3% gamma index passing rates of 94%–99%, validating our simulations with the MC model. The calculated organ/tissue doses per prescribed doses for the craniospinal irradiations (1 mGy Gy−1 to 1 Gy Gy−1) were generally much higher than those for the intracranial irradiations (2.1 μGy Gy−1 to 0.1 Gy Gy−1), which is due to the larger field coverage of the craniospinal irradiations. The largest difference was observed at the adrenal dose, i.e. ∼3000 times. In addition, the calculated organ/tissue doses were compared with those calculated with a simplified MC model, showing that the beam properties (i.e. spot size, spot divergence, mean energy, and energy spread) do not significantly influence dose calculations despite the limited irradiation cases. This implies that the use of the MC model commissioned to the MPTC measurement data might be dosimetrically acceptable for patient dose reconstructions at other proton centers particularly when their measurement data are unavailable. The developed MC model will be used to reconstruct organ/tissue doses for MPTC pediatric patients collected in the PPCR.

Radiation dosimetry is an essential input for epidemiological studies of radiotherapy patients aimed at quantifying the dose-response relationship of late-term morbidity and mortality. Individualised organ dose must be estimated for all tissues of interest located in-field, near-field, or out-of-field. Whereas conventional measurement approaches are limited to points in water or anthropomorphic phantoms, computational approaches using patient images or human phantoms offer greater flexibility and can provide more detailed three-dimensional dose information. In the current study, we systematically compared four different dose calculation algorithms so that dosimetrists and epidemiologists can better understand the advantages and limitations of the various approaches at their disposal. The four dose calculations algorithms considered were as follows: the (1) Analytical Anisotropic Algorithm (AAA) and (2) Acuros XB algorithm (Acuros XB), as implemented in the Eclipse treatment planning system (TPS); (3) a Monte Carlo radiation transport code, EGSnrc; and (4) an accelerated Monte Carlo code, the x-ray Voxel Monte Carlo (XVMC). The four algorithms were compared in terms of their accuracy and appropriateness in the context of dose reconstruction for epidemiological investigations. Accuracy in peripheral dose was evaluated first by benchmarking the calculated dose profiles against measurements in a homogeneous water phantom. Additional simulations in a heterogeneous cylinder phantom evaluated the performance of the algorithms in the presence of tissue heterogeneity. In general, we found that the algorithms contained within the commercial TPS (AAA and Acuros XB) were fast and accurate in-field or near-field, but not acceptable out-of-field. Therefore, the TPS is best suited for epidemiological studies involving large cohorts and where the organs of interest are located in-field or partially in-field. The EGSnrc and XVMC codes showed excellent agreement with measurements both in-field and out-of-field. The EGSnrc code was the most accurate dosimetry approach, but was too slow to be used for large-scale epidemiological cohorts. The XVMC code showed similar accuracy to EGSnrc, but was significantly faster, and thus epidemiological applications seem feasible, especially when the organs of interest reside far away from the field edge.

Epidemiological investigation is an important approach to assessing the risk of late effects after radiotherapy, and organ dosimetry is a crucial part of such analysis. Computed tomography (CT) images, if available, can be a valuable resource for individualizing the dosimetry, because they describe the specific anatomy of the patient. However, CT images acquired for radiation treatment planning purposes cover only a portion of the body near the target volume, whereas for epidemiology, the interest lies in the more distant normal tissues, which may be located outside the scan range. To address this challenge, we developed a novel method, called the Anatomically Predictive Extension (APE), to extend a partial-body CT image stack using images of a computational human phantom matched to the patient based on their height and weight. To test our method, we created five APE phantoms from chest and abdominal images extracted from the chest-abdomen-pelvis (CAP) CT scans of five patients. Organ doses were calculated for simple chest and prostate irradiations that were planned on the reference computational phantom (assumed patient geometry if no CT images are available), APE phantoms (patient-phantom hybrid given a partial-body patient CT) and full patient CAP CT scans (ground truth). The APE phantoms and patient CAP CT scans resulted in nearly identical dosimetry for those organs that were fully included in the partial-body CT used to construct the APE. The calculated doses to these same organs in the reference phantoms differed by up to 20% and 52% for the chest and prostate cases, respectively. For organs outside the scan coverage, the reference phantom showed, on average, dose differences of 31% (chest case) and 41% (prostate case). For the APE phantoms, these values were 26% (chest) and 17% (prostate). The APE method combines patient and phantom images to improve organ dosimetry both inside and outside the scan range. We intend to use the APE method for estimating dose for organs peripheral to the treatment fields; however, this method is quite generalizable with many potential applications.

No abstract

To develop an algorithm to automatically map CT scan locations of patients onto computational human phantoms to provide with patient-specific organ doses. We developed an algorithm that compares a two-dimensional skeletal mask generated from patient CTs with that of a whole body computational human phantom. The algorithm selected the scan locations showing the highest Dice Similarity Coefficient (DSC) calculated between the skeletal masks of a patient and a phantom. To test the performance of the algorithm, we randomly selected five sets of neck, chest, and abdominal CT images from the National Institutes of Health Clinical Center. We first automatically mapped scan locations of the CT images on a computational human phantom using our algorithm. We had several radiologists to manually map the same CT images on the phantom and compared the results with the automated mapping. Finally, organ doses for automated and manual mapping locations were calculated by an in-house CT dose calculator and compared to each other. The visual comparison showed excellent agreement between manual and automatic mapping locations for neck, chest, and abdomen-pelvis CTs. The difference in mapping locations averaged over the start and end in the five patients was less than 1 cm for all neck, chest, and AP scans: 0.9, 0.7, and 0.9 cm for neck, chest, and AP scans, respectively. Five cases out of ten in the neck scans show zero difference between the average manual and automatic mappings. Average of absolute dose differences between manual and automatic mappings was 2.3, 2.7, and 4.0% for neck, chest, and AP scans, respectively. The automatic mapping algorithm provided accurate scan locations and organ doses compared to manual mapping. The algorithm will be useful in cases requiring patient-specific organ dose for a large number of patients such as patient dose monitoring, clinical trials, and epidemiologic studies.

Active learning (AL) is a prominent technique for reducing the annotation effort required for training machine learning models. Deep learning offers a solution for several essential obstacles to deploying AL in practice but introduces many others. One of such problems is the excessive computational resources required to train an acquisition model and estimate its uncertainty on instances in the unlabeled pool. We propose two techniques that tackle this issue for text classification and tagging tasks, offering a substantial reduction of AL iteration duration and the computational overhead introduced by deep acquisition models in AL. We also demonstrate that our algorithm that leverages pseudo-labeling and distilled models overcomes one of the essential obstacles revealed previously in the literature. Namely, it was shown that due to differences between an acquisition model used to select instances during AL and a successor model trained on the labeled data, the benefits of AL can diminish. We show that our algorithm, despite using a smaller and faster acquisition model, is capable of training a more expressive successor model with higher performance. 1

Purpose: Recent advances in cancer treatments have greatly increased the likelihood of post‐treatment patient survival. Secondary malignancies, however, have become a growing concern. Epidemiological studies determining secondary effects in radiotherapy patients require assessment of organ‐specific dose both inside and outside the treatment field. An essential input for Monte Carlo modeling of particle transport is radiological images showing full patient anatomy. However, in retrospective studies it is typical to only have partial anatomy from CT scans used during treatment planning. In this study, we developed a multi‐step method to extend such limited patient anatomy to full body anatomy for estimating dose to normal tissues located outside the CT scan coverage. Methods: The first step identified a phantom from a library of body size‐dependent computational human phantoms by matching the height and weight of patients. Second, a Python algorithm matched the patient CT coverage location in relation to the whole body phantom. Third, an algorithm cut the whole body phantom and scaled them to match the size of the patient. Then, merged the two anatomies into one whole body. We entitled this new approach, Anatomically Predictive Extension (APE). Results: The APE method was examined by comparing the original chest‐abdomen‐pelvis CT images of the five patients with the APE phantoms developed from only the chest part of the CAP images and whole body phantoms. We achieved average percent differences of tissue volumes of 25.7%, 34.2%, 16.5%, 26.8%, and 31.6% with an average of 27% across all patients. Conclusion: Our APE method extends the limited CT patient anatomy to whole body anatomy by using image processing and computational human phantoms. Our ongoing work includes evaluating the accuracy of these APE phantoms by comparing normal tissue doses in the APE phantoms and doses calculated for the original full CAP images under generic radiotherapy simulations. This research was supported by the NIH Intramural Research Program.

No abstract