
==== Front
Phys Imaging Radiat Oncol
Phys Imaging Radiat Oncol
Physics and Imaging in Radiation Oncology
2405-6316
Elsevier

S2405-6316(24)00111-8
10.1016/j.phro.2024.100641
100641
Original Research Article
Automated plan generation for prostate radiotherapy patients using deep learning and scripted optimization
Church Cody cchurch@toh.ca
a⁎
Yap Michelle a
Bessrour Mohamed a
Lamey Michael a
Granville Dal b
a Department of Medical Physics, The Ottawa Hospital General Campus, Canada
b Department of Radiation Oncology and Department of Physics and Atmospheric Science, Dalhousie University, Canada
⁎ Corresponding author at: Department of Medical Physics, The Ottawa Hospital General Campus, 501 Smyth Rd, Ottawa, ON K1H 8L6, Canada. cchurch@toh.ca
08 9 2024
10 2024
08 9 2024
32 10064119 4 2024
30 8 2024
4 9 2024
© 2024 The Author(s)
2024
https://creativecommons.org/licenses/by-nc-nd/4.0/ This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
Background and Purpose

Treatment planning is a time-intensive task that could be automated. We aimed to develop a “single-click” workflow, fully deployed within a commercial treatment planning system (TPS), for autoplanning prostate radiotherapy treatment plans using predictions from a deep learning model (DLM).

Materials and Methods

Automatically generated treatment plans were created with a single script, executed from within a commercial TPS scripting environment, that performed two stages sequentially. Initially, a 3D dose distribution was predicted with a ResUNet DLM. The DLM was trained and validated using previously treated datasets (n = 120) which used 3D contours as inputs. Following this, dose predictions were converted into treatment plans by extracting dose-volume metrics from the predictions to use as objectives for the inverse optimizer within the TPS. An independent test dataset (n = 20) was used to evaluate the similarity between automated and clinical plans.

Results

For planning target volumes, the median percentage difference and interquartile range between the automatically generated plans and clinical plans were 0.4% [0.2-1.1%] for the V100%, −0.5% [(−1.0)-(−0.2)%] for D99% and −0.5% [(−1.0)-(−0.2)%] for D95%. Bladder and rectum volume-at-dose objectives agreed within −6.1% [(−12.5)-0.9%]. The conversion of the DLM prediction into a treatment plan took 15 min [13-16 min].

Conclusions

An automatic plan generation workflow that uses a DL model with scripted optimization was fully deployed in a commercial TPS. Autoplans were compared to previously treated clinical plans and were found to be non-inferior.

Keywords

Deep learning
Autoplanning
Radiotherapy
==== Body
pmc1 Introduction

Modern treatment planning in radiotherapy is commonly achieved with an inverse optimization process. Quality of treatment plans and the time taken to generate them can be affected by the skills and experience of the treatment planner [1]. As such, automating the inverse optimization process has the potential to improve efficiency and reduce variability [2], [3].

Recent studies in automated treatment planning typically employ a knowledge-based planning (KBP) approach. As summarized in Babier et al., KBP workflows involve a two-step process [4]. In the first step, the dose distribution for a new patient treatment is predicted using knowledge of previous, similar patient treatments. In the second step, a deliverable treatment plan (hereinafter shortened to treatment plan) is automatically generated to deliver a dose distribution that closely matches the prediction. We define a deliverable plan in the context of volumetric modulated arc therapy (VMAT) as a plan that respects the physical limitations of the linac (e.g. dose rates, multileaf collimator (MLC) motions, jaw motions, gantry angles, collimator angles, couch angles, etc.).

The first step (i.e. dose prediction) is often accomplished using machine learning [5], [6] or deep learning (DL). DL approaches may be preferable as they eliminate the need for feature engineering and can retain full three-dimensional (3D) dose information. Excellent performance of DL models for modulated dose distribution predictions has been demonstrated for a variety of treatment sites [7], [8], [9], [10], [11], [12], [13], [14]. Notably, all entries in the 2020 AAPM OpenKBP Grand Challenge used DL for dose prediction [15].

The second step of the KBP process (i.e. generation of a treatment plan that closely matches the prediction) is often achieved using dose mimicking algorithms [15]. Dose mimicking algorithms are not common in commercial treatment planning systems (TPSs), although one is available for fallback planning [16]. A limitation of this approach is that the predicted dose distribution may not be optimal, although this can potentially be addressed using objective functions that attempt to improve upon the predicted dose distribution, rather than simply matching it [17].

An alternative to dose mimicking is to use predicted dose distributions to generate objectives and weights that can be used in standard inverse optimization algorithms. Although less common in recent KBP studies, this approach has advantages. Inverse optimization tools are readily available in commercial TPSs and are widely used in radiotherapy workflows. Additionally, prominent commercial TPSs contain scripting tools that can be used to automate the process. This method may offer a more human-centered approach to automation as it allows planners to evaluate plans in familiar software and continue optimization in an intuitive manner, if desired.

A recent investigation examined the direct prediction of VMAT MLC patterns using DL, foregoing the optimization step [18]. This approach, though novel and promising, has thus far resulted in plans of insufficient quality, and requires “in-house” software to generate plans, which makes clinical deployment challenging.

Although the use of DL dose predictions in KBP pipelines has been well-demonstrated [4], [15], there are few studies that have used optimization algorithms within commercial TPSs to generate treatment plans. Using a commercial TPS for this step reduces barriers to clinical deployment as they are typically well-validated and approved by regulatory bodies. Xia et al. developed an automated method to generate rectum intensity-modulated radiation therapy (IMRT) plans that required exports from the TPS to an external DL server [19]. Lempart et al. derived optimization objectives using a nearest-neighbor search to identify the most similar dose distributions in an atlas of historical treatments [9]. This approach relied on availability of a robust atlas and included some manual steps. van de Sande et al. used a commercial dose mimicking algorithm designed for fallback planning to create breast IMRT plans based on two-dimensional (2D) DL dose predictions [20]. To our knowledge, this algorithm cannot be automated and requires manual intervention.

In this work, we developed a fully-automated workflow for generating VMAT treatment plans for prostate radiotherapy within a commercial TPS. To our knowledge, this is the first study demonstrating a fully-automated workflow that uses a custom DL dose prediction model and scripted optimization to automatically generate VMAT plans within a commercial TPS. Potential benefits of this approach include the use of common optimization objectives (which planners can intuitively interact with to tweak the plan), the elimination of data transfers, full deployment within a commercial TPS scripting environment, and improvements to dose distributions beyond mimicking predictions.

2 Materials and methods

2.1 Data description

Data from 140 previously treated prostate radiotherapy patients were used in this study. All patients were treated with dual-arc VMAT plans that delivered 60 Gy to the prostate and 54 Gy to the proximal seminal vesicles in 20 fractions. Each dataset included the 3D dose distribution and contours of the planning target volume (PTV) PTV60 (prostate + margins), PTV54 (proximal 1 cm of seminal vesicles + margins), bladder, rectum, left/right femurs (contoured inferiorly to the extent of the ischial tuberosities) and the external body contour. Patients with hip prostheses were not included. Treatment plans in this dataset were created in the Monaco TPS (Elekta AB, Stockholm) and dose was calculated using Monaco’s Monte Carlo (MC) dose engine. This study was approved by the Ottawa Health Science Network Research Ethics Board (ID#: 20200300 01H).

2.2 Deep learning dose prediction model

A subset of the dataset described in section 2.1 was used to train (n=100) and validate (n=20) a 3D residual U-Net to predict 3D dose distributions [21], [22]. Structure contours and dose data were extracted from digital imaging and communications in medicine (DICOM) files exported from Monaco and used to train the DL model. Structures were converted to binary masks and dose distributions were normalized such that D95% of the PTV60 was equal to 60 Gy, which is a common normalization method used in our clinic. Structure and dose arrays were resized and cropped to extend 72 mm inferiorly and 120 mm superiorly about the isocentre resulting in 64×256×256 arrays with 1.2×1.2×3 mm voxel size to match the computed tomography (CT) images. This cropping was chosen to balance the intent to minimize the dataset size (reducing computational memory burden) while maintaining the full extent of the key organs at risk (OARs) (e.g. rectum and bladder).

The first layer of the DL model used 8 filters and each subsequent layer used fi=2i∗f0, where f0 = 8. The model was trained with 1000 epochs (batch size = 1) and a learning rate of 0.0001 with the Adam optimizer. The model architecture is shown in Fig. 1. TensorFlow [23] and an NVIDIA RTX 8000 GPU were used for model training. The model was trained using a custom, domain-specific loss function inspired by Soomro et al. and Sun et al. [24], [25]. This loss function summed a voxel-wise weighted mean absolute error (MAE) function with absolute dose-volume histogram (DVH) metric differences between predicted and delivered dose distributions. Weights in the MAE calculation varied by structure and dose level. Voxels within targets (i.e. PTV60 and PTV54) and key OARs (i.e. bladder and rectum) were assigned higher weights, while voxels in low-dose areas and less important structures were assigned lower weights. The weight parameters were tuned based upon the expertise of two independent observers to preserve the dose distributions expected with local plans using the validation dataset. A visual depiction of the loss function is included in Fig. 1. DVH metrics used in the loss calculation were derived from internal institutional protocols based on the CHHiP trial [26], [27] and represent important metrics in local plan-quality evaluation. These are summarized in Table 1.Fig. 1 A schematic representation of the residual U-Net architecture used in this work. The numbers above each block represent the spatial dimensions and number of feature channels along each convolutional layer. Dashed lines with green squares represent dense connections (concatenation) between the encoder and decoder regions. Skip connections (addition) are represented by red circles. The output of the network is a 3D dose distribution with the same spatial resolution as the input structures. The customized loss function is visualized in the bottom left which features a summation of the voxel-weighted mean absolute error (with the relative weightings shown) with a summation of all the difference between dose-volume metrics.

Table 1 Dose volume metrics that were included in the custom loss function. In this table, DX refers to the dose received by X% structure volume and VY refers the volume receiving Y Gy.

Structure	Dose Volume Metrics	
PTV60	D99%, D95%, D10%	
PTV54	D99%, D95%, D10%	
Bladder	V60Gy, V48.6Gy, V40.8Gy, Dmax	
Rectum	V60Gy, V57Gy, V52.8Gy, V48.6Gy, V40.8Gy, V32.4Gy, V24.6Gy, Dmax	
FemurRight	V36Gy	
FemurLeft	V36Gy	

2.3 Scripted optimization

The RayStation (RaySearch Laboratories, Stockholm) Python scripting application programming interface (API) was used to automatically optimize treatment plans (autoplans) using DVH metrics extracted from the predicted dose distributions. This scripted routine was developed by using simple DVH-based optimization objectives in RayStation to mimic (and, if possible, improve upon) DVH curves from clinical plans (created in Monaco) in the validation dataset (n=20). Once the routine was able to reproduce the clinical plans adequately, it was then used to produce plans that match predicted dose distributions in the autoplanning pipeline.

The DVH-based optimization objectives and weights used in this routine were chosen for their simplicity and were adjusted through a trial-and-error approach. Because there is not a simple, consensus metric to determine similarity or superiority of dose distributions, we relied on the expertise of two independent observers to inform adjustments to the objectives to improve the routine. This process was analogous to common treatment planning workflows, in which planners first start with ideal optimization objectives and adjust until they receive an adequate result based on their clinical expertise.

The DVH-based objectives used in this routine are summarized in Table 2. Because predicted dose distributions are not necessarily optimal, we attempted to improve upon them by setting DVH objectives that were slightly better than those predicted. For all OARs, the predicted D1.0% and Dmax values were reduced by 1.0% to create ‘MaxDVH’ and ‘MaxDose’ objectives. Similarly, the volumes predicted to receive a given dose had the doses reduced by 1.0 Gy to create ‘MaxDVH’ objectives (e.g. the predicted V60Gy was used to create a V59Gy ‘MaxDVH’ objective). For target volumes (i.e. PTV54 and PTV60), the predicted DVH metrics that quantify coverage (i.e. V54gy, V59Gy, V60Gy) were increased by 1.0% to a maximum of 100% of the volume to create ‘MinDVH’ objectives. The predicted DVH metrics that quantify hot spots (i.e. V55.5Gy, V57Gy, D1.0%, and Dmax) were reduced by 1.0% to create ‘MaxDVH’ and ‘MaxDose’ objectives. In addition to the DVH objectives, a single ‘DoseFallOff’ objective was used to control conformality.Table 2 Dose volume metrics that were calculated from the dose predicted by the deep learning model. These metrics were used as dose-volume objectives in the automated TPS optimization process. Organs at risk include the rectum, bladder, and individual femurs.

Structure	Dose Volume Metrics	
PTV60	V60Gy, V59Gy, D1.0%, Dmax	
PTV54	V54Gy, V55.5Gy, V57Gy, D1.0%, Dmax	
Organs At Risk	V60Gy, V50Gy, V40Gy, V30Gy, V20Gy, V10Gy, D1.0%, Dmax	

For the initial optimization, all objectives were given a weight of 1. After each iteration, objectives that were not met had their weight increased by 50. This process was repeated for 10 iterations. The choice of 10 iterations was found to balance improvements in the plan with total optimization time.

2.4 Fully automated planning deployment

The autoplanning process described in 2.2, 2.3 was fully deployed within the RayStation Python scripting API. A Python virtual environment was created in RayStation, and the packages necessary for executing the DL dose prediction model were installed (e.g. TensorFlow and its dependencies). Execution of a single script within RayStation resulted in an autoplan being generated according to the following automated steps (also shown in Fig. 2). Initially, the patient’s 3D contours were converted into binary arrays for input into the DL model described in section 2.2 to generate a predicted 3D dose distribution. DVH metrics were then extracted from the predicted dose distribution and used to create a patient-specific optimization template. Following this, the VMAT treatment plan characteristics were initialized (e.g. energy, arc geometry, couch, and collimator positions) and the iterative optimization process described in section 2.3 was executed to create a treatment plan.Fig. 2 The workflow for automatic treatment plan generation using the RayStation scripting API. The native Python language and built-in functions within this environment allow for dose prediction with a DL model. Dose-volume metrics from the predicted dose were used to create and execute a patient-specific optimization and generate a plan with the built-in TPS optimizer.

2.5 Statistical evaluation and analysis

An independent test dataset (n=20) was used to evaluate the full autoplanning process. A two-sided Wilcoxon signed-rank test with a significance level of 0.05 was performed to compare DVH metrics used to assess plan quality in our institution (outlined in Fig. 3) between the autoplan and the clinical plan. To adjust for multiple comparisons, a Bonferroni correction was used with n=18 comparisons to adjust the significance level to 0.0028. This correction is for the comparisons between the metrics shown in Fig. 3 (metrics not shown for femurs as they were equal to zero) and the conformity index [28] for the PTV60 and PTV54. All metrics are reported as the median and interquartile range.Fig. 3 Comparison of various dose-volume metrics between the deep learning dose prediction (Prediction), the deliverable plan generated with the automated workflow (Autoplan), and the previously delivered clinical plan (Clinical). The centre tick of the bars represents the median value across all patients (n = 20) and the box represents the upper and lower quartile range. The left-axis references all dosimetric metrics and the right axis references all volumetric metrics.

3 Results

The conversion of the prediction into the autoplan took 15 min [13-16 min] across the test set. A comparison of the DVH-based plan-quality metrics used in our local institutional protocols is shown for the DL dose prediction, autoplan, and clinical plan in Fig. 3. With respect to these metrics, n=15 patients passed all PTV metrics in the clinical plans with n=5 cases exceeding the primary threshold for maximum dose (63 Gy). For the autoplans, all patients passed all metrics for the PTVs. There was no significant difference between the conformity index found for the PTV60 (clinical: 0.92 [0.91-0.94], autoplan: 0.93 [0.91-0.93]) or the PTV54 (clinical: 0.20 [0.17-0.26], autoplan: 0.22 [0.18-0.27]). With respect to metrics for OARs, 91.5% were met across all clinical plans and 91.9% were met across all autoplans. The plan-quality metrics that most frequently failed were V60 for the bladder and rectum. Failure rates for the bladder V60 were 55% in both the clinical and autoplans. Failure rates for the rectum V60 were 45% for the clinical plans and 40% for the autoplans. Between the autoplan and the clinical plan, statistically significant differences were seen for V60 and D1.0% in PTV60. Significant differences were seen for V54 in PTV54. For OARs, significant differences were only seen in the bladder V48.6 (p=4.8e-04) and V40.8 (p=7.1e-04). The most notable differences between the clinical and autoplans were hotter doses received by PTV54 (increase of 6% [3-13%] in volume receiving 57 Gy) and smaller volumes receiving doses in the ∼12–20 Gy range.

4 Discussion

In this work, a fully automated, “single-click” planning process deployed within a commercial TPS was presented. This process produced prostate VMAT plans that were non-inferior to previous clinically approved plans. There are several benefits to this automated process: A treatment plan is generated with a “single-click.” The model used for predicting dose is potentially sharable and does not contain any personal health information. Specialized hardware/software is not needed (beyond the TPS). The scripting solution is designed with freely available software (Python). The treatment plans are created within the confines of a validated commercial TPS. The final plan and optimization template is in a familiar format that planners can further improve and/or tweak. No data transfers are required.

In the first stage of autoplan generation, a 3D dose distribution was predicted with a residual U-Net trained with 3D contours and previous clinical dose distributions. A potential limitation of this work is the combination of model architecture and loss function selected. Rigorous testing of multiple architecture and loss function pairings was not explored and is beyond the scope of this work. Our goal was to generate a model with sufficient performance for our purposes, rather than rigorously comparing multiple modelling methods. There is no clear consensus on the “best” choice of architecture and loss function for dose prediction models, though many combinations have been studied [4], [7], [8], [10], [11], [15], [20], [29], [30], [31]. While the speed of model training could be impacted by the model parameters and data configuration, this step is likely inconsequential from a clinical-utility perspective; once a model is trained, calling it to predict a 3D dose, even on a CPU, should only take seconds to a few minutes with modern hardware.

Strategies to automatically generate treatment plans from DL dose predictions within clinical TPSs have varied in the literature. Lempart et al. [9] derived optimization objectives using a nearest neighbor search that compared predictions to an atlas of previously treated cases. This work required the availability of a sizeable and robust atlas and relied on manual intervention for DICOM transfers and setting of some optimization objectives. In the RayStation TPS, van de Sande et al. [20] created deliverable doses for left-sided breast cancer patients using a commercial dose-mimicking algorithm on doses predicted by a 2D U-Net model. They found statistically significant differences in the mean dose received by the PTV – although the magnitude of the difference may not be clinically significant. Xia et al. [19] performed autoplanning using a similar strategy to this work, in that DL dose predictions were used to generate optimization objectives in a clinical TPS in a fully-automated manner. It differed, however, in that autoplans had a single target (rectum), less complex treatment modality (IMRT), used a different TPS (Pinnacle, Philips Radiation Oncology Systems), and required DICOM data transfers to a dedicated DL server.

One limitation of this study is the unavailability of clinical plans developed in RayStation to compare to the autoplans. RayStation was used to generate autoplans due to its mature Python scripting API, but available clinical plans were generated using Monaco. Dose distributions in Monaco were calculated using MC, while RayStation used collapsed cone superposition-convolution. Because of the statistical uncertainty in MC calculations, there are larger maximum doses [32], [33] and less-steep DVHs. The magnitude of the impact is larger in smaller volumes [34]. A characteristic example comparing an autoplan and clinical plan is shown in Fig. 4. As seen in this example, the DVHs for PTV60 and PTV54 exhibit a more homogenous dose in the autoplan. This observation is generally consistent across all patients in the independent test set and is at least partially attributable to differences in dose engines and beam models, rather than genuine differences in plan quality. Similarly, we could not compare manually generated RayStation plans to autoplans. However, we have no reason to expect manually generated RayStation plans to substantially differ from the Monaco plans, outside of the differing dose engines. Favourable comparisons with the clinically approved Monaco plans suggest that the autoplans are of sufficient quality to be deemed clinically acceptable.Fig. 4 A characteristic example comparing a clinical plan (previously treated) to automatically generated plan derived from a dose predicted by the deep learning model. (A) Shows the dose volume histograms for all relevant structures where the solid lines are from the clinical plan and the dashed lines are from the autoplan. Subfigures (B), (D) and (F) shown an axial, coronal and sagittal slice from the clinical dose distribution overlayed on the CT. Subfigures (C), (E) and (G) shown an axial, coronal and sagittal slice from the autoplan dose distribution overlayed on the CT.

In conclusion, a single-click automated plan generation routine fully deployed within the scripting environment of a commercial TPS was presented. Automated plans were non-inferior to manually generated plans when evaluated using common DVH-based metrics.

Financial disclosures

This work was supported by the Harold E. Johns Scholarship.

Data availability statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.

CRediT authorship contribution statement

Cody Church: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writing – original draft, Writing – review & editing. Michelle Yap: Methodology, Software, Validation. Mohamed Bessrour: Methodology, Software. Michael Lamey: Conceptualization, Supervision, Funding acquisition, Writing – review & editing. Dal Granville: Conceptualization, Data curation, Supervision, Funding acquisition, Methodology, Writing – review & editing.

Declaration of competing interest

The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: Cody Church – No declaration. Michael Lamey – No declaration. Michelle Yap - Funded by the Harold E. Johns Scholarship. Mohamed Bessrour - Funded by the Harold E. Johns Scholarship. Dal Granville - Currently Employed by Nova Scotia Health Authority in the Department of Radiation Oncology and appointed with the Department of Physics and Atmospheric Science, Dalhousie University.
==== Refs
References

1 Nelms B.E. Robinson G. Markham J. Velasco K. Boyd S. Narayan S. Variation in external beam treatment plan quality: an inter-institutional study of planners and planning systems Pract Radiat Oncol 2 2012 296 305 10.1016/j.prro.2011.11.012 24674168
2 Scaggion A. Fusella M. Roggio A. Bacco S. Pivato N. Rossato M.A. Reducing inter-and intra-planner variability in radiotherapy plan output with a commercial knowledge-based planning solution Phys Med 53 2018 86 93 10.1016/j.ejmp.2018.08.016 30241759
3 van Gysen K. Kneebone A. Le A. Wu K. Haworth A. Bromley R. Evaluating the utility of knowledge-based planning for clinical trials using the TROG 08.03 post prostatectomy radiation therapy planning data Phys Imaging Radiat Oncol 22 2022 91 97 10.1016/j.phro.2022.05.004 35602546
4 Babier A. Mahmood R. Zhang B. Alves V.G. Barragán-Montero A.M. Beaudry J. OpenKBP-Opt: an international and reproducible evaluation of 76 knowledge-based planning pipelines Phys Med Biol 67 2022 185012 10.1088/1361-6560/ac8044
5 McIntosh C. Welch M. McNiven A. Jaffray D.A. Purdie T.G. Fully automated treatment planning for head and neck radiotherapy using a voxel-based dose prediction and dose mimicking method Phys Med Biol 62 2017 5926 10.1088/1361-6560/aa71f8 28486217
6 Shiraishi S. Moore K.L. Knowledge-based prediction of three-dimensional dose distributions for external beam radiotherapy Med Phys 43 2016 378 387 10.1118/1.4938583 26745931
7 Nguyen D. Jia X. Sher D. Lin M.-H. Iqbal Z. Liu H. 3D radiotherapy dose prediction on head and neck cancer patients with a hierarchically densely connected U-net deep learning architecture Phys Med Biol 64 2019 065020 10.1088/1361-6560/ab039b
8 Mahmood R, Babier A, McNiven A, Diamant A, Chan TC. Automated treatment planning in radiation therapy using generative adversarial networks. PMLR: Proceedings of the 3rd machine learning for healthcare conference; 2018; Palo Alto, California. California: 2018; p. 484–499.
9 Lempart M. Benedek H. Gustafsson C.J. Nilsson M. Eliasson N. Bäck S. Volumetric modulated arc therapy dose prediction and deliverable treatment plan generation for prostate cancer patients using a densely connected deep learning model Phys Imaging Radiat Oncol 19 2021 112 119 10.1016/j.phro.2021.07.008 34401537
10 Bakx N. Bluemink H. Hagelaar E. van der Sangen M. Theuws J. Hurkmans C. Development and evaluation of radiotherapy deep learning dose prediction models for breast cancer Phys Imaging Radiat Oncol 17 2021 65 70 10.1016/j.phro.2021.01.006 33898781
11 Ahn S.H. Kim E. Kim C. Cheon W. Kim M. Lee S.B. Deep learning method for prediction of patient-specific dose distribution in breast cancer Radiat Oncol 16 2021 1 13 10.1186/s13014-021-01864-9 33402192
12 Koike Y. Takegawa H. Anetai Y. Ohira S. Nakamura S. Tanigawa N. Patient-specific three-dimensional dose distribution prediction via deep learning for prostate cancer therapy: improvement with the structure loss Phys Med 107 2023 102544 10.1016/j.ejmp.2023.102544
13 Murakami Y. Magome T. Matsumoto K. Sato T. Yoshioka Y. Oguchi M. Fully automated dose prediction using generative adversarial networks in prostate cancer patients PloS one 15 2020 e0232697 32365088
14 Nguyen D. Long T. Jia X. Lu W. Gu X. Iqbal Z. A feasibility study for predicting optimal radiation therapy dose distributions of prostate cancer patients from patient anatomy using deep learning Sci Rep 9 2019 1076 10.1038/s41598-018-37741-x 30705354
15 Babier A. Zhang B. Mahmood R. Moore K.L. Purdie T.G. McNiven A.L. OpenKBP: the open-access knowledge-based planning grand challenge and dataset Med Phys 48 2021 5549 5561 34156719
16 RayStation B.D. External beam treatment planning system Med Dosim 43 2018 168 176 29650302
17 Yousefi A. Ketabi S. Abedi I. A novel mathematical model to generate semi-automated optimal IMRT treatment plan based on predicted 3D dose distribution and prescribed dose Med Phys 50 2023 3148 3158 10.1002/mp.16236 36691067
18 Heilemann G. Zimmermann L. Schotola R. Lechner W. Peer M. Widder J. Generating deliverable DICOM RT treatment plans for prostate VMAT by predicting MLC motion sequences with an encoder-decoder network Med Phys 50 2023 5088 5094 10.1002/mp.16545 37314944
19 Xia X. Wang J. Li Y. Peng J. Fan J. Zhang J. An artificial intelligence-based full-process solution for radiotherapy: a proof of concept study on rectal cancer Front Oncol 10 2021 616721 10.3389/fonc.2020.616721
20 van de Sande D. Sharabiani M. Bluemink H. Kneepkens E. Bakx N. Hagelaar E. Artificial intelligence based treatment planning of radiotherapy for locally advanced breast cancer Phys Imaging Radiat Oncol 20 2021 111 116 10.1016/j.phro.2021.11.007 34917779
21 Han X. Automatic liver lesion segmentation using a deep convolutional neural network method arXiv:170407239 2017. https://doi.org/10.48550/arXiv.1704.07239.
22 Milletari F, Navab N, Ahmadi S-A. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In: IEEE: Proc. 2016 Fourth International Conference on 3D Vision (3DV); 2016; Stanford, California. California: 2016; p. 565-571.
23 Chollet F. Deep learning with Python 2021 Simon and Schuster
24 Soomro MH, Alves VGL, Nourzadeh H, Siebers JV. DeepDoseNet: a deep learning model for 3D dose prediction in radiation therapy arXiv preprint arXiv:211100077 2021. https://doi.org/10.48550/arXiv.2111.00077.
25 Sun Z. Xia X. Fan J. Zhao J. Zhang K. Wang J. A hybrid optimization strategy for deliverable intensity-modulated radiotherapy plan generation using deep learning-based dose prediction Med Phys 49 2022 1344 1356 10.1002/mp.15462 35043971
26 Khoo V. Dearnaley D. Question of dose, fractionation and technique: ingredients for testing hypofractionation in prostate cancer–the CHHiP trial Clin Oncol 20 2008 12 14 10.1016/j.clon.2007.10.008
27 Morgan S.C. Hoffman K. Loblaw D.A. Buyyounouski M.K. Patton C. Barocas D. Hypofractionated radiation therapy for localized prostate cancer: an ASTRO, ASCO, and AUA evidence-based guideline J Clin Oncol 36 2018 3411 10.1016/j.prro.2018.08.002
28 Paddick I. Lippitz B. A simple dose gradient measurement tool to complement the conformity index J Neurosurg 105 2006 194 205 10.3171/sup.2006.105.7.194
29 Babier A. Mahmood R. McNiven A.L. Diamant A. Chan T.C. Knowledge-based automated planning with three-dimensional generative adversarial networks Med Phys 47 2020 297 306 10.1002/mp.13896 31675444
30 Cardenas C.E. McCarroll R.E. Court L.E. Elgohari B.A. Elhalawani H. Fuller C.D. Deep learning algorithm for auto-delineation of high-risk oropharyngeal clinical target volumes with built-in dice similarity coefficient parameter optimization function Int J Radiat Oncol Biol Phys 101 2018 468 478 10.1016/j.ijrobp.2018.01.114 29559291
31 Gronberg M.P. Gay S.S. Netherton T.J. Rhee D.J. Court L.E. Cardenas C.E. Dose prediction for head and neck radiotherapy using a three-dimensional dense dilated U-net architecture Med Phys 48 2021 5567 5573 10.1002/mp.14827 34157138
32 Bukulmez T. Ozdemir B. The Effect of Pencil Beam, Collapsed Cone, and Monte Carlo Algorithms on Dose-Volume Parameters in Esophagus Cancer The Digital Phantom Study Turk J Oncol 36 2021 191 199 10.5505/tjo.2021.2517
33 Hasenbalg F, Neuenschwander H, Mini R, Born E. Collapsed cone and analytical anisotropic algorithm dose calculations compared to VMC++ Monte Carlo simulations in clinical cases. IOP Publishing: J Phys Conf Ser; 2007; Het Pand, Gent, Belgium. Belgium: 2007; p. 021007.
34 Feygelman V. Latifi K. Bowers M. Greco K. Moros E.G. Isacson M. Maintaining dosimetric quality when switching to a Monte Carlo dose engine for head and neck volumetric-modulated arc therapy planning J Appl Clin Med Phys 23 2022 e13572 35213089
