PaperlayerView exact revision on arXiv

Source-linked AI summary

Knowledge-based automated planning with three-dimensional generative adversarial networks

Aaron Babier, Rafid Mahmood, Andrea L. McNiven, Adam Diamant, Timothy C. Y. Chan

arXiv:1812.09309v1physics.med-ph

TL;DR

Conventional treatment planning can take several days, while earlier neural-network KBP methods predicted slice-wise doses. This paper builds a 3D GAN-based planning pipeline and reports better clinical-criteria performance and more realistic dose predictions than previous state-of-the-art KBAP.

  • Problem

    Conventional radiation planning may take several days, motivating automated approaches that could reduce plan-generation lead time.

    Source: §1 · pp. 1–2

  • Method

    The pipeline predicts dose with 2D slice-wise or full-volume 3D GAN variants, then estimates objective weights and optimizes fluence-based plans, also testing prediction scaling.

    Source: §1 · p. 2, §2.1 · p. 4, §2.3 · p. 5

  • Results

    The 3D GAN predicted smoother, more realistic dose distributions across the longitudinal axis, and the final framework outperformed previous state-of-the-art KBAP in satisfying clinical criteria.

    Source: §4 · pp. 10, 12, §5 · p. 13

  • Takeaways & Limitations

    The framework generated high-quality treatments without manual intervention in the evaluated oropharyngeal planning setting.

    Source: §5 · p. 13

    GAN predictions lack interpretability, making it harder for planners to understand when and why prediction errors occur.

    Source: §4 · p. 13

Abstract

from arXiv · show

We develop a knowledge-based automated planning (KBAP) pipeline that generates treatment plans using deep neural network architectures for predicting 3D doses. Our pipeline consisted of a generative adversarial network (GAN) to predict dose from a CT image followed by two optimization models to learn objective function weights and generate fluence-based plans, respectively. We investigated three different GAN models. The first two models predicted dose for each axial slice independently. One predicted dose as a RGB color map, while the other predicted a scalar value for dose directly. The third GAN model predicted scalar doses for the full 3D CT image at once, considering correlations between adjacent CT slices. For all models, we also investigated the impact of scaling the GAN predictions before optimization. Each GAN model was trained on 130 previously delivered oropharyngeal treatment plans. Performance was tested on 87 out-of-sample plans, by evaluating using clinical planning criteria and compared to their corresponding clinical plans. The best performing KBAP plans were generated with the 3D GAN, which predicted dose values followed by scaling. These plans satisfied close to 77% of all clinical criteria, compared to the clinical plans, which satisfied 64% of all criteria. Additionally, these KBAP plans satisfied the same criteria as the clinical plans 84% more frequently compared to the 2D GAN model using RGB dose prediction. The 3D GAN predictions were also more similar to the final plan compared to the other approaches, as it better captured the vertical dosimetric relationship between adjacent axial slices. The deliverable plans better captured implicit constraints associated with physical deliverability. Overall, our final plans had superior performance in satisfying clinical criteria and generated more realistic predictions compared to the previous state-of-the-art.

1 Introduction

Conventional radiation planning can take several days, motivating automated approaches that could shorten plan-generation lead time. This paper develops a 3D GAN-based KBP method and tests its modifications within an oropharynx KBAP pipeline.

1 Introduction

  • Conventional radiation planning uses iterative exchanges between planners and oncologists and may take several days; automation could reduce plan-generation lead time.The paper notes potential value for producing high-quality plans for complicated sites and serving growing radiation-therapy demand.Source: §1 · pp. 1–2
  • KBAP first predicts a clinically acceptable dose, then uses optimization to convert it into a deliverable plan.The prediction component learns dose characteristics from historical treatment plans, and final plan quality strongly correlates with prediction quality.Source: §1 · p. 2
  • Earlier neural-network KBP methods predicted 2D dose distributions for axial CT slices and concatenated them into a 3D distribution.Related approaches included convolutional, generative adversarial, and recurrent neural networks.Source: §1 · p. 2
  • The paper develops the first 3D GAN-based KBP method, predicting a full 3D dose distribution from a 3D image at once.It also investigates scalar dose prediction and scaling before optimization.Source: §1 · p. 2
  • The tested KBAP modifications achieved better out-of-sample performance than state-of-the-art benchmarks.The experiments varied full 3D dose prediction, scalar voxel-dose output, and scaling predictions before optimization.Source: §1 · p. 2

2 Methods and Materials

The methods train three GAN variants on historical oropharynx plans, then use their dose predictions in a two-stage optimization pipeline. The study evaluates clinical-criteria satisfaction and prediction similarity across out-of-sample plans.

2.1 Prediction Using Generative Adversarial Networks

  • Conditional GAN training combines adversarial terms with an l1 loss that makes generated dose distributions resemble clinical doses.The generator receives Gaussian input and a CT image; the discriminator judges whether a dose distribution is clinically delivered, and λ balances adversarial loss against resemblance to deliverable plans.Source: §2.1 · pp. 3–4
  • The GAN variants predict dose either slice by slice as RGB or scalar values, or for the full 3D image as scalar values.The 3D-dose model uses 3D convolution and deconvolution filters; the two 2D variants concatenate predicted axial slices.Source: §2.1 · p. 4

2.2 Training the GAN

  • The dataset comprised 217 single-institution oropharyngeal treatment plans, split into 130 training plans and 87 out-of-sample test plans.The plans used 6 MV step-and-shoot IMRT; CT images were converted to 128 axial slices of 128 × 128 voxels.Source: §2.2 · p. 5

2.3 Creating Plans Using Inverse Planning

  • A two-stage inverse-planning pipeline estimates objective weights from each predicted dose, then optimizes fluence-based plans under an SPG limit of 55.SPG is used as a convex surrogate for physical deliverability; scaled variants multiply predictions by the smallest amount needed to satisfy all target dose criteria before optimization.Source: §2.3 · pp. 5–6

2.4 Performance Analysis

  • Plan quality was assessed by the fraction of clinical criteria met and by comparing KBAP plans with clinical plans across OARs, targets, and all ROIs.Table 2 defines the planning criteria, including D99, Dmean, and Dmax.Source: §2.4 · p. 6, Table 2 · p. 7
  • Prediction quality was assessed using average absolute error between predicted-dose and KBAP-plan DVHs for each ROI and patient.The analysis compared 2D-RGB with 2D-dose, 2D-dose with scaled 2D-dose, and scaled 2D-dose with scaled 3D-dose.Source: §2.4 · p. 6

3 Results

Scaled 3D-dose′ plans most closely matched clinical plans’ pass/fail pattern, while scaled 2D-dose′ and 3D-dose′ plans satisfied the highest fractions of all ROI criteria. In prediction comparisons, 3D-dose′ had lower median error than 2D-dose′ and 2D-RGB, and 3D predictions better captured dose relationships across adjacent slices.

KBAP Plan Quality

  • 3D-dose′ plans agreed most with clinical plans on passed and failed criteria; where they differed, they satisfied five times as many criteria as clinical plans.For all ROI criteria, 2D-dose′ and 3D-dose′ satisfied 77.0% and 76.4%, respectively.Source: KBAP Plan Quality · p. 7, Table 3 · p. 8
  • Among criteria also passed clinically, 3D-dose and 3D-dose′ performed best across OAR and target criteria, with 3D-dose′ highest for targets and most OARs.3D-dose′ was not highest for larynx or mandible; multiple approaches tied for top results on brainstem, esophagus, and PTV63.Source: KBAP Plan Quality · p. 7, Table 4 · p. 8
  • 78.2% of 3D-dose′ plans satisfied all criteria met by their corresponding clinical plans, the highest proportion among approaches.The largest scaling-related improvement in satisfying all ROI criteria was 34.5 percentage points between 3D-dose and 3D-dose′; 3D-dose′’s stronger target coverage came at the expense of OAR sparing relative to 3D-dose.Source: KBAP Plan Quality · pp. 7–8, Table 5 · p. 9

KBP Prediction Quality

  • 3D-dose′ median prediction error was 0.07 Gy lower than 2D-dose′ and 0.63 Gy smaller than 2D-RGB.Scalar dose encoding reduced median error by 0.19 Gy versus RGB; scaling reduced half of 2D-dose predictions’ errors by about 0.31 Gy, while PTV56 was less accurate after scaling (average median error of 1.00 Gy).Source: KBP Prediction Quality · p. 9
  • The 3D GAN better learned vertical relationships between adjacent axial slices, producing smoother dose predictions than the streaky discontinuities in 2D predictions.The 2D dose falloff could be impossibly steep near target boundaries, whereas 3D predictions had more realistic gradients.Source: KBP Prediction Quality · p. 9

4 Discussion

The discussion highlights three KBAP enhancements—scalar dose prediction, scaling predictions before optimization, and full 3D dose prediction—and reports improved plan quality. It also notes the pipeline’s clinical-planning rationale and limitations.

  • The approach produced plans better than the previous state-of-the-art in KBAP while generating full dose distributions rather than summary statistics.Source: §4 · p. 10
  • The framework combines scalar voxel-dose prediction, prediction scaling before optimization, and full 3D dose prediction from 3D CT images.Source: §4 · p. 10
  • 74% more often, 2D-dose′ plans satisfied the same clinical-plan criteria across all ROIs than 2D-RGB′ plans.The authors suggest scalar prediction may be easier than predicting an RGB triplet.Source: §4 · p. 10
  • 52% more often, scaled plans satisfied the same criteria as clinical plans than unscaled plans; scaled plans also satisfied 10% more criteria overall.Source: §4 · p. 10
  • Scaling is automatic and may correct small inaccuracies in absolute dose, while the GAN appears better at learning dose variation across tissues than exact tissue doses.Source: §4 · p. 11
  • The reported criterion goal reflects that only 68.4% of criteria were achieved across 217 clinically delivered plans, and the authors also valued matching each plan’s original criteria.Source: §4 · p. 12
  • The authors chose DVH-based inverse planning after dose prediction because it aligns with clinical practice; they also avoid hand-tailored feature engineering.They suggest institutions could train GANs on plans reflecting their own guidelines.Source: §4 · p. 12
  • The approach separates prediction from optimization, and GAN predictions lack interpretability, making it harder for planners to understand prediction errors.Source: §4 · pp. 12–13

5 Conclusion

The paper presents a 3D GAN-based KBAP framework and reports superior clinical-criteria performance and more realistic predictions than the previous state-of-the-art.

  • Using 217 oropharyngeal cancer treatment plans, the framework’s one-channel prediction, post-prediction scaling, and full 3D dose prediction supported high-quality treatment plans without manual intervention.Source: §5 · p. 13

Related papers

Loading 1812.09309v1…