Are age and gender suitable matching criteria in organ dose reconstruction using surrogate childhood cancer patients’ CT scans?

Abstract

Purpose: The purpose of this work was to assess the feasibility of using surrogate CT scans of matched patients for organ dose reconstructions for childhood cancer (CC) survivors, treated in the past with only 2D imaging data available instead of 3D CT data, and in particular using the current literature standard of matching patients based on similarity in age and gender. Methods: Thirty-one recently treated CC patients with abdominal CT scans were divided into six age-and gender-matched groups. From each group, two radiotherapy plans for Wilms' tumor were selected as reference plans and applied to the age-and gender-matched patients' CTs in the respective group. Two reconstruction strategies were investigated: S1) without field adjustments; S2) with manual field adjustments according to anatomical information, using a visual check in digitally reconstructed radiographs. To assess the level of agreement between the reconstructed and the reference dose distributions, we computed (using a collapsed cone algorithm) and compared the absolute deviation in mean and maximum dose normalized by the prescribed dose (i.e., normalized errors |NE mean | and |NE 2cc |) in eight organs at risk (OARs): heart, lungs, liver, spleen, kidneys, and spinal cord. Furthermore, we assessed the quality of a reconstruction case by varying acceptance thresholds for |NE mean | and |NE 2cc |. A reconstruction case was accepted (i.e., considered to pass) if the errors in all OARs are smaller than the threshold. The pass fraction for a given threshold was then defined as the percentage of reconstruction cases that were classified as a pass. Furthermore, we consider the impact of allowing to use a different CT scan for each OAR. Results: Slightly smaller reconstruction errors were achieved with S2 in multiple OARs than with S1 (P < 0.05). Among OARs, the best reconstruction was found for the spinal cord (average |NE mean | and |NE 2cc | ≤ 4%). The largest average |NE mean | was found in the spleen (18%). The largest average |NE 2cc | was found in the left lung (26%). Less than 30% of the reconstruction cases (i.e., pass fraction) meet the criteria that |NE mean | < 20% and |NE 2cc | < 20% in all OARs when using age and gender matching and a single CT to do reconstructions. Allowing other matchings and combining reconstructions for OARs from multiple patients, the pass fraction increases substantially to more than 60%. Conclusions: To conclude, reconstructions with small deviations can be obtained by using CC patients' CT scans, making the general approach promising. However, using age and gender as the only matching criteria to select a CT scan for the reconstruction is not sufficient to guarantee sufficiently low reconstruction errors. It is therefore suggested to include more features (e.g., height, features extracted from 2D radiographs) than only age and gender for dose reconstruction for CC survivors treated in the pre-3D radiotherapy planning era and to consider ways to combine multiple reconstructions focused on different OARs.

Machine learning for the prediction of pseudorealistic pediatric abdominal phantoms for radiation dose reconstruction

“…Figure 9 shows the effect of using GP-GOMEA (the overall best performing algorithm) when learning ML models of the liver position using different training set sizes in a traditional fivefold cross validation. The test set size consists of 1 5 th of the total number of patients (60∕5 ¼ 12 patients), whereas the maximum number of patients to be considered at training time is varied between 1 5 th (12 patients) and 4 5 th (48 patients). Patients are picked at random (taking care that patients in the test set never appear in the training set).…”

“…2,3,10,17 Since the lack of resemblance between the patient's and the phantom's anatomy is a primary source of error, 18 this can ultimately lead to inaccurate dose reconstructions. 4 We remark that, for the purpose of dose reconstruction, anatomy resemblance (configuration and properties of internal organs) does not necessarily imply anatomy realism. 8 To improve phantom individualization for patients with no 3-D imaging, adaptation techniques have been studied such as morphing organ shapes 19,20 and organ repositioning.…”

“…1,2 We consider the scenario of radiation dose reconstruction for childhood cancer survivors who were treated in the pre-3-D planning era, i.e., before computed tomography (CT) scans became commonplace in radiation treatment practice. 3,4 Reconstructing the radiation dose that these patients were subject to is key to study the possible late adverse effects of radiation treatment and consequently improve future treatments. [5][6][7] Here phantoms are needed to act as a surrogate for the unknown 3-D anatomy of the patient.…”

Automatic radiotherapy plan emulation for 3D dose reconstruction to enable big data analysis for historically treated patients

“…The plans were visualized on DRRs in Oncentra TPS. The manual plan emulation strategy is described in [22] and the resulting plans were approved by a radiation oncologist. The planner was not provided with information on which plan was manually or automatically emulated.…”

Machine learning for automatic construction of pediatric abdominal phantoms for radiation dose reconstruction

“…As no 3D imaging is available for historical patients to simulate the treatment and estimate the radiation dose distribution, phantoms are necessary to act as surrogate anatomies in order to reconstruct 3D dose distributions. 12,13 We focus on children between 2 to 6 years, and on dose reconstruction for abdominal radiation treatment, for the following reasons. Firstly, children are typically under-represented in existing phantom libraries, i.e., phantoms are available for few categories.…”