A Deep Learning-based correction to EPID dosimetry for attenuation and scatter in the Unity MR-Linac system

Comparison of deep learning models for building two-dimensional non-transit EPID Dosimetry on Varian Halcyon

Background

This study compared the effectiveness of five deep learning models in constructing non-transit dosimetry with an a-Si electronic portal imaging device (EPID) on Varian Halcyon. Deep learning model is increasingly used to support prediction and decision-making in several fields including oncology and radiotherapy.

Materials and methods

Forty-seven unique plans of data obtained from breast cancer patients were calculated using Eclipse treatment planning system (TPS) and extracted from DICOM format as the ground truth. Varian Halcyon was then used to irradiate the a-Si 1200 EPID detector without an attenuator. The EPID and TPS images were augmented and divided randomly into two groups of equal sizes to distinguish the validation and training-test data. Five different deep learning models were then created and validated using a gamma index of 3%/3 mm.

Results

Four models successfully improved the similarity of the EPID images and the TPS-generated planned dose images. Meanwhile, the mismatch of the constituent components and number of parameters could cause the models to produce wrong results. The average gamma pass rates were 90.07 ± 4.96% for A-model, 77.42 ± 7.18% for B-model, 79.60 ± 6.56% for C-model, 80.21 ± 5.88% for D-model, and 80.47 ± 5.98% for E-model.

Conclusion

The deep learning model is proven to run fast and can increase the similarity of EPID images with TPS images to build non-transit dosimetry. However, more cases are needed to validate this model before being used in clinical activities.

Deep learning-based tools to distinguish plan-specific from generic deviations in EPID-based in vivo dosimetry

BACKGROUND

Dose distributions calculated with electronic portal imaging device (EPID)-based in vivo dosimetry (EIVD) differ from planned dose distributions due to generic and plan-specific deviations. Generic deviations are characteristic to a class of plans. Examples include limitations in EIVD dose reconstruction, inaccuracies in treatment planning system (TPS) calculations and systematic machine deviations. Plan-specific deviations have an unpredictable character. Examples include discrepancies between the patient model used for dose calculation and the patient position or anatomy during delivery, random machine deviations, and data transfer, human or software errors. During the inspection work performed with traditional γ-evaluation statistical methods: (i) generic deviations raise alerts that need to be inspected but that rarely lead to action as their root cause is usually understood and (ii) the detection of relevant plan-specific deviations may be hindered by the presence of generic deviations.

PURPOSE

To investigate whether deep learning-based tools can help in identifying γ-alerts raised by generic deviations and in improving the detectability of plan-specific deviations.

METHODS

A 3D U-Net was trained as an autoencoder to reconstruct underlying patterns of generic deviations in γ-distributions. The network was trained for four treatment disease sites differently affected by generic deviations: volumetric modulated arc therapy (VMAT) lung (no known deviations), VMAT prostate (TPS inaccuracies), VMAT head-and-neck (EIVD limitations) and intensity modulated radiation therapy (IMRT) breast (large EIVD limitations). The network was trained with virtual non-transit γ-distributions: 60 train/10 validation for the VMAT sites and 30 train/10 validation for IMRT breast. It was hypothesized that in vivo γ-distributions obtained in the presence of plan-specific deviations would differ from those seen during training. For each disease site, the sensitivity of γ-analysis and the network to detect (synthetically introduced) patient-related deviations was compared by receiver operator characteristic analysis. The investigated deviations were patient positioning errors, weight gain or loss, and tumor volume changes. The clinical relevance was illustrated qualitatively with 793 in vivo clinical cases (141 lung, 136 head-and-neck, 209 prostate and 307 breast).

RESULTS

Error detectability of patient-related deviations was better with the network than with γ-analysis. The average area under the curve values over all sites were 0.86 ± 0.12(1SD) and 0.69 ± 0.25(1SD), respectively. Regarding in vivo clinical results, the percentage of cases differently classified by γ-analysis and the network was 1%, 19%, 18% and 64% for lung, head-and-neck, prostate, and breast, respectively. In head-and-neck and breast cases, 45 γ-only alerts were examined, of which 43 were attributed to EPID dose reconstruction limitations. For prostate, all 15 investigated γ-only alerts were due to known TPS inaccuracies. All 59 investigated network alerts were explained by either patient-related deviations or EPID acquisition incidents. Some patient-related deviations detected by the network were not detected by γ-analysis.

CONCLUSIONS

Deep learning-based tools trained to reconstruct underlying patterns of generic deviations in γ-distributions can be used to (i) automatically identify false positives within the set of γ-alerts and (ii) improve the detection of plan-specific deviations, hence minimizing the likelihood of false negatives. The presented method provides clear additional value to the γ-alert management process for large scale EIVD systems.

Towards real-time EPID-based 3D in vivo dosimetry for IMRT with Deep Neural Networks: A feasibility study

We investigate the potential of the Deep Dose Estimate (DDE) neural network to predict 3D dose distributions inside patients with Monte Carlo (MC) accuracy, based on transmitted EPID signals and patient CTs. The network was trained using as input patient CTs and first-order dose approximations (FOD). Accurate dose distributions (ADD) simulated with MC were given as training targets. 83 pelvic CTs were used to simulate ADDs and respective EPID signals for subfields of prostate IMRT plans (gantry at 0∘). FODs were produced as backprojections from the EPID signals. 581 ADD-FOD sets were produced and divided into training and test sets. An additional dataset simulated with gantry at 90∘ (lateral set) was used for evaluating the performance of the DDE at different beam directions. The quality of the FODs and DDE-predicted dose distributions (DDEP) with respect to ADDs, from the test and lateral sets, was evaluated with gamma analysis (3%,2 mm). The passing rates between FODs and ADDs were as low as 46%, while for DDEPs the passing rates were above 97% for the test set. Meaningful improvements were also observed for the lateral set. The high passing rates for DDEPs indicate that the DDE is able to convert FODs into ADDs. Moreover, the trained DDE predicts the dose inside a patient CT within 0.6 s/subfield (GPU), in contrast to 14 h needed for MC (CPU-cluster). 3D in vivo dose distributions due to clinical patient irradiation can be obtained within seconds, with MC-like accuracy, potentially paving the way towards real-time EPID-based in vivo dosimetry.

A convolutional neural network model for EPID-based non-transit dosimetry

PURPOSE

To develop an alternative computational approach for EPID-based non-transit dosimetry using a convolutional neural network model.

METHOD

A U-net followed by a non-trainable layer named True Dose Modulation recovering the spatialized information was developed. The model was trained on 186 Intensity-Modulated Radiation Therapy Step & Shot beams from 36 treatment plans of different tumor locations to convert grayscale portal images into planar absolute dose distributions. Input data were acquired from an amorphous-Silicon Electronic Portal Image Device and a 6 MV X-ray beam. Ground truths were computed from a conventional kernel-based dose algorithm. The model was trained by a two-step learning process and validated through a five-fold cross-validation procedure with sets of training and validation of 80% and 20%, respectively. A study regarding the dependance of the amount of training data was conducted. The performance of the model was evaluated from a quantitative analysis based the ϒ-index, absolute and relative errors computed between the inferred dose distributions and ground truths for six square and 29 clinical beams from seven treatment plans. These results were also compared to those of an existing portal image-to-dose conversion algorithm.

RESULTS

For the clinical beams, averages of ϒ-index and ϒ-passing rate (2%-2mm > 10% D_max ) of 0.24 (±0.04) and 99.29 (±0.70)% were obtained. For the same metrics and criteria, averages of 0.31 (±0.16) and 98.83 (±2.40)% were obtained with the six square beams. Overall, the developed model performed better than the existing analytical method. The study also showed that sufficient model accuracy can be achieved with the amount of training samples used.

CONCLUSION

A deep learning-based model was developed to convert portal images into absolute dose distributions. The accuracy obtained shows that this method has great potential for EPID-based non-transit dosimetry.

Automatic dosimetric verification of online adapted plans on the Unity MR-Linac using 3D EPID dosimetry

BACKGROUND AND PURPOSE

The Unity MR-Linac is equipped with an EPID, the images from which contain information about the dose delivered to the patient. The purpose of this study was to introduce a framework for the automatic dosimetric verification of online adapted plans using 3D EPID dosimetry and to present the obtained dosimetric results.

MATERIALS AND METHODS

The framework was active during the delivery of 1207 online adapted plans corresponding to 127 clinical IMRT treatments (74 prostate, 19 rectum, 19 liver and 15 lymph node oligometastases). EPID reconstructed dose distributions in the patient geometry were calculated automatically and then compared to the dose distributions calculated online by the treatment planning system (TPS). The comparison was performed by γ-analysis (3% global/2mm/10% threshold) and by the difference in median dose to the high-dose volume (ΔHDV_D50). 85% for γ-pass rate and 5% for ΔHDV_D50 were used as tolerance limit values.

RESULTS

93% of the online plans were verified automatically by the framework. Missing EPID data was the reason for automation failure. 91% of the verified plans were within tolerance.

CONCLUSION

Automatic dosimetric verification of online adapted plans on the Unity MR-Linac is feasible using in vivo 3D EPID dosimetry. Almost all online adapted plans were approved automatically by the framework. This newly developed framework is a major step forward towards the clinical implementation of a permanent safety net for the entire online adaptive workflow.

Feasibility of using a commercial collapsed cone dose engine for 1.5T MR-LINAC online independent dose verification

PURPOSE

To validate the feasibility and accuracy of commonly used collapsed cone (CC) dose engine for Elekta Unity 1.5T MR-LINAC online independent dose verification.

MATERIALS AND METHODS

The Unity beam model was built and commissioned in RayStation treatment planning system with CC dose engine. Four AAPM TG-119 test plans were created and measured with ArcCHECK phantom for comparison, another thirty patient plans from six tumor sites were also included. The dosimetric criteria for various ROIs and 3D gamma passing rates were quantitatively evaluated, and the effects of magnetic field and dose deposition type on the dose difference between two systems were further analyzed.

RESULTS

ArcCHECK based measurement showed a clear magnetic field induced profile shift between CC with both measurement and GPUMCD. For clinical plans, gamma passing rates with criteria (3%, 3 mm) between GPUMCD and CC large than 90% can be achieved for most tumor sites except esophagus and lung cases, the mean dose difference of 3% can be satisfied for most ROIs from all tumor sites. The magnetic field caused a large dose impact on low density areas, the average gamma passing rates were improved from 85.54% to 96.43% and 87.40% to 99.54% for esophagus and lung cases when the magnetic field effect was excluded.

CONCLUSIONS

It is feasible to use CC dose engine as a secondary dose calculation tool for Elekta Unity system for most tumor sites, while the accuracy is limited and should be used carefully for low density areas, such as esophagus and lung cases.