PaperlayerView exact revision on arXiv

Source-linked AI summary

Knowledge-based Radiation Treatment Planning: A Data-driven Method Survey

Shadab Momin, Yabo Fu, Yang Lei, Justin Roper, Jeffrey D. Bradley, Walter J. Curran, Tian Liu, Xiaofeng Yang

arXiv:2009.07388v2physics.med-ph

TL;DR

Manual intervention in inverse treatment planning can take up to a few days, motivating data-driven dose prediction from prior plans. This review surveys traditional and deep-learning-based KBP, reporting that traditional methods can match or exceed experienced planners while recent deep-learning methods show greater potential.

  • Problem

    Inverse treatment planning can require extensive manual intervention to achieve a high-quality dose distribution, taking up to a few days before treatment begins.

    Source: §1 · p. 2

  • Method

    The review classifies data-driven dose prediction as traditional KBP using geometric or anatomical features, or deep-learning-based methods training neural networks to predict dose.

    Source: §1 · p. 2

  • Results

    Traditional KBP approaches are reported as equivalent or superior to an experienced planner with greater efficiency, while recent deep-learning developments hold greater potential for dose prediction.

    Source: §5 · p. 22

  • Takeaways & Limitations

    Both KBP approaches need expansion to more complex sites, including abdominal and intracranial cases.

    Source: §5 · p. 22

    The review covers external-beam radiation therapy and excludes brachytherapy and patient-specific treatment-plan quality-assurance prediction.

    Source: §2.2 · p. 3

Abstract

from arXiv · show

This paper surveys the data-driven dose prediction approaches introduced for knowledge-based planning (KBP) in the last decade. These methods were classified into two major categories according to their methods and techniques of utilizing previous knowledge: traditional KBP methods and deep-learning-based methods. Previous studies that required geometric or anatomical features to either find the best matched case(s) from repository of previously delivered treatment plans or build prediction models were included in traditional methods category, whereas deep-learning-based methods included studies that trained neural networks to make dose prediction. A comprehensive review of each category is presented, highlighting key parameters, methods, and their outlooks in terms of dose prediction over the years. We separated the cited works according to the framework and cancer site in each category. Finally, we briefly discuss the performance of both traditional KBP methods and deep-learning-based methods, and future trends of both data-driven KBP approaches.

1 Introduction

Inverse treatment planning can require substantial manual intervention and take up to a few days to produce a high-quality plan. This review compares traditional and deep-learning-based knowledge-based planning (KBP), which use prior plans differently to predict dose for new patients.

1 Introduction

  • Inverse planning can require substantial manual intervention and take up to a few days before a patient receives the first radiation fraction.Data-driven planning aims to use knowledge from previous plans to predict outcomes for new patients.Source: §1 · p. 2
  • Traditional KBP uses geometric features to retrieve similar plans or build prediction models, whereas deep-learning methods learn patterns from raw data without manual feature extraction.The review groups atlas-based, statistical-modeling, and machine-learning approaches as traditional KBP.Source: §1 · p. 2
  • The review covers methods, parameters, dose-prediction investigations, performance, challenges, and future trends across both KBP categories.Source: §1 · p. 2

2 Methods

The review searched multiple scholarly databases and screened peer-reviewed articles for methods and comparable or improved dose-prediction quality or efficiency. It defined the included treatment-planning scope and summarized publication trends for traditional and deep-learning-based approaches.

2.1 Literature search

  • The literature search used Scopus, Web of Science, PubMed, Google Scholar, and arXiv’s medical-physics category with KBP, machine-learning, deep-learning, and dose-prediction terms.Source: §2.1 · p. 3

2.2 Article selection criteria

  • The review included peer-reviewed external-beam studies with described methods and comparable or improved dose-prediction quality or efficiency, including retrospective RapidPlan studies.Brachytherapy and patient-specific treatment-plan quality-assurance prediction were excluded.Source: §2.2 · p. 3
  • Here, dose prediction spans DVHs, dose metrics, voxel or spatial distributions, plan objectives or constraints, and transferring these to a new case to generate a plan.Source: §2.2 · p. 3
  • Deep-learning dose-prediction publications rose from 1 in 2016 to 15 in 2020, while traditional KBP publication counts increased between 2015 and 2018.Figure 1 presents annual and cumulative publication counts for both categories.Source: §2.2 · p. 3

3 Results

This section reviews data-driven KBP dose-prediction methods, covering traditional atlas- and model-based approaches and deep-learning methods across a range of prediction tasks, models, and learning strategies.

3.1 Knowledge Based Planning

  • Traditional KBP includes over 90 articles classified as atlas-based or model-based, predicting entire DVHs, dose metrics, or voxel doses.Atlas methods match prior cases using physical or anatomical similarity and transfer planning knowledge; model-based methods use handcrafted geometric features with statistical or machine-learning regression.Source: §3.1 · pp. 4–5
  • OVH-based pancreatic predictions were within 1 and 2 Gy for more than 82% and 94% of patients, with 1.4 Gy and 1.7 Gy improvements in mean liver and kidney doses, respectively.OVH was also found to outperform hydrogel volume for representing rectum sparing; incorporating spatial information improved prediction for some sites but not breast cavity or lung.Source: §3.1 · pp. 6–7
  • Generalized DTH improved DVH prediction for multiple head-and-neck organs and was extended to select plans using a weighted similarity metric across OARs.DTH is equivalent to OVH when Euclidean distance is used; multivariate regression with DTH is commercially available as RapidPlanTM.Source: §3.1 · pp. 8–9
  • Dosimetric outliers degraded prediction, and inconsistent planning priorities increased errors; sample-size needs varied by model, endpoint, and method.About 30 plans were sufficient for dose-volume predictions with less than 3% relative error in head-and-neck and whole-pelvis/prostate; atlas methods required larger cohorts than ANN or SVR.Source: §3.1 · pp. 9–10

3.2 Deep Learning

  • Deep-learning dose prediction includes ANN and DNN methods, with CNN architectures widely used alongside GANs and a treatment-planning application of reinforcement learning.CNNs use convolution, pooling, and related components to extract features; U-Net supports image-to-image prediction, while reinforcement learning trains a virtual treatment planner network to adjust planning parameters.Source: §3.2 · pp. 12–13

3.3 Deep Learning in Dose Prediction

  • Nearly 30 DL-based dose-prediction publications were reported, with prostate and head-and-neck among the most studied sites; U-Net was widely used to predict dose distributions.Early applications included 2D dose prediction, including predicting Acuros XB dose from lower-resolution AAA dose and CT.Source: §3.3 · p. 14
  • HD U-Net outperformed standard U-Net for 3D dose prediction, while Xing et al. mapped ray-tracing dose to collapsed-cone-calculated dose [156].HD U-Net combines DenseNet feature propagation with U-Net local and global feature inference; the ray-tracing input omits scatter effects.Source: §3.3 · p. 15
  • For Pareto-dose prediction, Ma et al. used 3D U-Net, Nguyen et al. added DVH-based and adversarial losses, and Bohara et al. incorporated beam information.The added losses address clinically relevant differences between dose distributions for the same anatomy.Source: §3.3 · p. 15
  • U-Net models predicted internal-radiation 3D dose-rate maps, while ResNets were used for head/neck IMRT and nasopharyngeal Helical Tomotherapy.The internal-dose studies used mass-density and radioactivity maps, aiming for precision close to Monte Carlo simulations; ResNet shortcut connections mitigate vanishing gradients.Source: §3.3 · p. 15

3.3.2 Overview of GAN-based works

  • DoseGAN outperformed conventional 2D and 3D GANs across all dosimetric criteria, including PTV and OARs.The first 3D GAN predicted full 3D dose distributions and outperformed the earlier 2D GAN, possibly by learning relationships between adjacent axial slices.Source: §3.3 · p. 15

3.3.3 Overview of learning processes

  • Unsupervised learning was the most widely used strategy for DL dose prediction; supervised learning used labeled targets, while semi-supervised learning addressed partially available targets.GANs are a semi-supervised approach used for dose prediction.Source: §3.3 · p. 16

4 Discussion

The discussion contrasts traditional and deep-learning KBP, reviews their comparative dose-prediction results, and identifies data, outlier, and deliverability challenges alongside future automation directions.

4 Discussion

  • Traditional KBP publications exceeded 90 between 2011 and August 2020, while more than 15 DL dose-prediction papers had appeared in 2020 by August.Prostate, head/neck, and lung were commonly investigated; direct sample-size comparisons were not made because DL studies were fewer per site.Source: §4 · p. 20
  • DVH prediction covers contoured OARs but may miss conformity and hotspots outside them; voxel prediction addresses this scope issue but is vulnerable to outlier plans.The review reports no well-established workflow for identifying and mitigating outliers in either KBP approach.Source: §4 · p. 20
  • Traditional KBP relies on substantial feature extraction, while DL can model nonlinear patterns; GANs may suit heterogeneous-dose tasks but are difficult to train.The review anticipates network extensions such as attention-gated GANs to address conventional GAN shortcomings.Source: §4 · p. 21

4.1 Method comparison of KBP dose predictions

  • Across cited comparisons, DL methods showed stronger dose-prediction performance, including better parotid-dose coverage, GAN clinical-plan similarity, and attention-gated GAN performance for SBRT hotspots.The review attributes DL's edge presumably to its ability to use more than the small feature sets of traditional KBP.Source: §4.1 · p. 21

4.2 Future trends

  • Traditional KBP work is expected to continue using commercial tools such as RapidPlan, while DL research is shifting toward adaptive radiotherapy and broader planning tasks.Only one reported study had adopted DL for head/neck adaptive radiotherapy at the time of the review.Source: §4.2 · pp. 21–22
  • Low dose-prediction error does not guarantee a clinically deliverable plan; the discussion favors linking prediction with inverse optimization or direct fluence-to-plan generation.The review also points to patient-specific quality-assurance prediction and transfer learning as future workflow directions.Source: §4.2 · p. 22

5 Conclusion

The conclusion reviews traditional and newer DL-based KBP dose-prediction approaches, noting their potential and the need to extend them toward complex sites and automated planning.

5 Conclusion

  • The review covers over 90 traditional dose-prediction articles and nearly 30 more recently introduced DL-based KBP articles.It describes traditional approaches as equivalent or superior to an experienced planner with greater efficiency, and DL as holding greater potential for dose prediction.Source: §5 · p. 22
  • Both KBP approaches need expansion to more complex sites, including abdominal and intracranial cases.The conclusion also anticipates more retrospective studies using commercially accessible RapidPlan.Source: §5 · p. 22
  • The long-term goal identified by the review is a fully automated treatment-planning system.Source: §5 · p. 22

Related papers

Loading 2009.07388v2…