Multi-institutional generalizability of a plan complexity machine learning model for predicting pre-treatment quality assurance results in radiotherapy.

Background and purpose: Treatment plans in radiotherapy are subject to measurement-based pre-treatment verifications. In this study, plan complexity metrics (PCMs) were calculated per beam and used as input features to develop a predictive model. The aim of this study was to determine the robustness against differences in machine type and institutional-specific quality assurance (QA). Material and methods: A number of 567 beams were collected, where 477 passed and 90 failed the pre-treatment QA. Treatment plans of different anatomical regions were included. One type of linear accelerator was represented. For all beams, 16 PCMs were calculated. A random forest classifier was trained to distinct between acceptable and non-acceptable beams. The model was validated on other datasets to investigate its robustness. Firstly, plans for another machine type from the same institution were evaluated. Secondly, an inter-institutional validation was conducted on three datasets from different centres with their associated QA. Results: Intra-institutionally, the PCMs beam modulation, mean MLC gap, Q1 gap, and Modulation Complexity Score were the most informative to detect failing beams. Eighty-tree percent of the failed beams (15/18) were detected correctly. The model could not detect over-modulated beams of another machine type. Inter-institutionally, the model performance reached higher accuracy for centres with comparable equipment both for treatment and QA as the local institute. Conclusions: The study demonstrates that the robustness decreases when major differences appear in the QA platform or in planning strategies, but that it is feasible to extrapolate institutional-specific trained models between centres with similar clinical practice. Predictive models should be developed for each machine type.

Machine learning-based treatment couch parameter prediction in support of surface guided radiation therapy.

Purpose: A fully independent, machine learning-based automatic treatment couch parameters prediction was developed to support surface guided radiation therapy (SGRT)-based patient positioning protocols. Additionally, this approach also acts as a quality assurance tool for patient positioning. Materials/Methods: Setup data of 183 patients, divided into four different groups based on used setup devices, was used to calculate the difference between the predicted and the acquired treatment couch value. Results: Couch parameters can be predicted with high precision µ = 0.90 , σ = 0.92 . A significant difference (p < 0.01) between the variances of Lung and Brain patients was found. Outliers were not related to the prediction accuracy, but are due to inconsistencies during initial patient setup. Conclusion: Couch parameters can be predicted with high accuracy and can be used as starting point for SGRT-based patient positioning. In case of large deviations (>1.5 cm), patient setup has to be verified to optimally use the surface scanning system.

Machine learning-based detection of aberrant deep learning segmentations of target and organs at risk for prostate radiotherapy using a secondary segmentation algorithm.

Objective.The output of a deep learning (DL) auto-segmentation application should be reviewed, corrected if needed and approved before being used clinically. This verification procedure is labour-intensive, time-consuming and user-dependent, which potentially leads to significant errors with impact on the overall treatment quality. Additionally, when the time needed to correct auto-segmentations approaches the time to delineate target and organs at risk from scratch, the usability of the DL model can be questioned. Therefore, an automated quality assurance framework was developed with the aim to detect in advance aberrant auto-segmentations.Approach. Five organs (prostate, bladder, anorectum, femoral head left and right) were auto-delineated on CT acquisitions for 48 prostate patients by an in-house trained primary DL model. An experienced radiation oncologist assessed the correctness of the model output and categorised the auto-segmentations into two classes whether minor or major adaptations were needed. Subsequently, an independent, secondary DL model was implemented to delineate the same structures as the primary model. Quantitative comparison metrics were calculated using both models' segmentations and used as input features for a machine learning classification model to predict the output quality of the primary model.Main results. For every organ, the approach of independent validation by the secondary model was able to detect primary auto-segmentations that needed major adaptation with high sensitivity (recall = 1) based on the calculated quantitative metrics. The surface DSC and APL were found to be the most indicated parameters in comparison to standard quantitative metrics for the time needed to adapt auto-segmentations.Significance. This proposed method includes a proof of concept for the use of an independent DL segmentation model in combination with a ML classifier to improve time saving during QA of auto-segmentations. The integration of such system into current automatic segmentation pipelines can increase the efficiency of the radiotherapy contouring workflow.

RealDRR - Rendering of realistic digitally reconstructed radiographs using locally trained image-to-image translation.

INTRODUCTION: Digitally reconstructed radiographs (DRRs) represent valuable patient-specific pre-treatment training data for tumor tracking algorithms. However, using current rendering methods, the similarity of the DRRs to real X-ray images is limited, requires time-consuming measurements and/or are computationally expensive. In this study we present RealDRR, a novel framework for highly realistic and computationally efficient DRR rendering. MATERIALS AND METHODS: RealDRR consists of two components applied sequentially to render a DRR. First, a raytracer is applied for forward projection from 3D CT data to a 2D image. Second, a conditional Generative Adverserial Network (cGAN) is applied to translate the 2D forward projection to a realistic 2D DRR. The planning CT and CBCT projections from a CIRS thorax phantom and 6 radiotherapy patients (3 prostate, 3 brain) were split in training and test sets for evaluating the intra-patient, inter-patient and inter-anatomical region generalization performance of the trained framework. Several image similarity metrics, as well as a verification based on template matching, were used between the rendered DRRs and respective CBCT projections in the test sets, and results were compared to those of a current state-of-the-art DRR rendering method. RESULTS: When trained on 800 CBCT projection images from two patients and tested on a third unseen patient from either anatomical region, RealDRR outperformed the current state-of-the-art with statistical significance on all metrics (two-sample t-test, p < 0.05). Once trained, the framework is able to render 100 highly realistic DRRs in under two minutes. CONCLUSION: A novel framework for realistic and efficient DRR rendering was proposed. As the framework requires a reasonable amount of computational resources, the internal parameters can be tailored to imaging systems and protocols through on-site training on retrospective imaging data.