ABSTRACT
PURPOSE: To create and investigate a novel, clinical decision-support system using machine learning (ML). METHODS AND MATERIALS: The ML model was developed based on 79 radiotherapy plans of brain tumor patients that were prescribed a total dose of 60 Gy delivered with volumetric-modulated arc therapy (VMAT). Structures considered for analysis included planning target volume (PTV), brainstem, cochleae, and optic chiasm. The model aimed to classify the target variable that included class-0 corresponding to plans for which the PTV treatment planning objective was met and class-1 that was associated with plans for which the PTV objective was not met due to the priority trade-off to meet one or more organs-at-risk constraints. Several models were evaluated using double-nested cross-validation and an area-under-the-curve (AUC) metric, with the highest performing one selected for further investigation. The model predictions were explained with Shapely additive explanation (SHAP) interaction values. RESULTS: The highest-performing model was Logistic Regression achieving an accuracy of 93.8 ± 4.1% and AUC of 0.98 ± 0.02 on the testing data. The SHAP analysis indicated that the ΔD99% metric for PTV had the greatest influence on the model predictions. The least important feature was ΔDMAX for the left and right cochleae. CONCLUSIONS: The trained model achieved satisfactory accuracy and can be used by medical physicists in a data-driven quality assurance program as well as by radiation oncologists to support their decision-making process in terms of treatment plan approval and potential plan modifications. Model explanation analysis showed that the model relies on clinically valid logic when making predictions.
ABSTRACT
In vivo dosimetry methods can verify the prescription dose is delivered to the patient during treatment. Unfortunately, in exit dosimetry, the megavoltage image is contaminated with patient-generated scattered photons. However, estimation and removal of the effect of this fluence improves accuracy of in vivo dosimetry methods. This work develops a 'tri-hybrid' algorithm combining analytical, Monte Carlo (MC) and pencil-beam scatter kernel methods to provide accurate estimates of the total patient-generated scattered photon fluence entering the MV imager. For the multiply-scattered photon fluence, a modified MC simulation method was applied, using only a few histories. From each second- and higher-order interaction site in the simulation, energy fluence entering all pixels of the imager was calculated using analytical methods. For photon fluence generated by electron interactions in the patient (i.e. bremsstrahlung and positron annihilation), a convolution/superposition approach was employed using pencil-beam scatter fluence kernels as a function of patient thickness and air gap distance, superposed on the incident fluence distribution. The total patient-scattered photon fluence entering the imager was compared with a corresponding full MC simulation (EGSnrc) for several test cases. These included three geometric phantoms (water, half-water/half-lung, computed tomography thorax) using monoenergetic (1.5, 5.5 and 12.5 MeV) and polyenergetic (6 and 18 MV) photon beams, 10 × 10 cm2 field, source-to-surface distance 100 cm, source-to-imager distance 150 cm and 40 × 40 cm2 imager. The proposed tri-hybrid method is demonstrated to agree well with full MC simulation, with the average fluence differences and standard deviations found to be within 0.5% and 1%, respectively, for test cases examined here. The method, as implemented here with a single CPU (non-parallelized), takes â¼80 s, which is considerably shorter compared to full MC simulation (â¼30 h). This is a promising method for fast yet accurate calculation of patient-scattered fluence at the imaging plane for in vivo dosimetry applications.
