Role of AI in RT Planning Process

Knowledge Based and AI assisted Planning

Dr Santam Chakraborty
drsantam@gmail.com

Tata Medical Center, Kolkata

Introduction

  • Conceptually a process of “programming” a treatment machine to deliver a certain dose of radiotherapy to the patient.
  • Simple concept -> complex execution
  • Parameters that need to be optimized1:
    • Time when radiation generator is “active”
    • Direction from where it is aimed and position of the patient
    • Aperture shape and dimensions through which beam comes out
    • Modulation of beam intensity across time
    • Integration of anatomical variations during the treatment

Evaluating Treatment Plans

  • The best treatment plan is often difficult to judge based on simple metrics
  • DVH parameters are not indicative of the anatomical complexity in the plan.
  • Factors which predict dosimetric parameters in a treatment plan:
    • Target volume
    • Target shape
    • Location of the target
    • Percentage of organ overlapping PTV
    • Distance to target / overlap volume histogram
    • Beam orientation
    • Distance between organs

Plan quality variation

  • Prospective study in prostate cancer with 14 defined plan quality metrics

  • 125 plans received from planners

  • Evaluated against pre-defined plan quality metrics

  • Plan quality widely varied BUT did not differ by:

    • TPS

    • Modality: VMAT / IMRT

    • Education/ certification of planner

  • Likely related to “planner skill”

Nelms et al (2012)Nelms BE, Robinson G, Markham J, Velasco K, Boyd S, Narayan S, et al. Variation in external beam treatment plan quality: An inter-institutional study of planners and planning systems. Pract Radiat Oncol 2012;2:296–305. https://doi.org/10.1016/j.prro.2011.11.012.

Dose Prediction

flowchart TD
   A[Dose Estimation Models] --> B[Heuristic based methods]
   A --> C[Library based methods]
   A --> D[Model based methods]
Figure 1
Property Heuristic Based Library Based Model Based
Principle An empiricial way to determine the dose fall off around target Use previous treated patients data to predict dose to next person Generate a model (ML) to predict dose of next patient
Application Organ dose obtained from the model of dose fall off around target Library of matched patients used to derive achievable plans Model is generated and predicts from new patient
Requirement Beam parameters Previous treated patients Model based on other patients

PlanIQ (Heuristic Based)

  • Calculates a “benchmark dose” - 3D dose grid with 100% dose to each target volume. 2

  • Estimates minimum dose that each voxel outside the target must receive to ensure this coverage

  • The dose spread is determined by the energy of the beam using a lookup table of dose distributions for a specific energy

  • A high dose spread is calculated along with a low dose spread and these two are “convolved” to generate a final dose grid.

  • The system generates DVHs that can be “achieved” for a given patient.

PlanIQ output figure from Perumal et al. Perumal B, Sundaresan HE, Ranganathan V, Ramar N, Anto GJ, Meher SR. Evaluation of plan quality improvements in PlanIQ-guided Autoplanning. Rep Pract Oncol Radiother 2019;24:533–43. https://doi.org/10.1016/j.rpor.2019.08.003.

PlanIQ output figure from Perumal et al. Perumal B, Sundaresan HE, Ranganathan V, Ramar N, Anto GJ, Meher SR. Evaluation of plan quality improvements in PlanIQ-guided Autoplanning. Rep Pract Oncol Radiother 2019;24:533–43. https://doi.org/10.1016/j.rpor.2019.08.003.

Library of Plan Method

  • ICON P Project conceived to test the efficacy of a library based plan evaluation system.

  • Cohort of 94 prostate cancer patients with approved plans treated with VMAT or HT (60 Gy / 20 fractions)

  • Volumes of PTV, rectum bladder and overlap structures with these volumes created.

  • A library of planned DVH was created where user can input the volumetric metrics, and the system would display a “subset” of plans which matched these metrics.

ICON P Methodology - Taken from Saha et al. 3

ICON P Study Design

ICON P Study Design (Saha et al)

Plan Constraints

Results

  • At least one target ideal or acceptable constraints achievable for each of rectum and bladder in 87.5%

  • Volume receiving 53 Gy in bladder reduced by 5.5% (SE: 0.92%)

  • Volume receiving 53 Gy in rectum reduced by 6.87% (SE: 1.04%)

Knowledge Based Planning

  • Using existing knowledge to predict the achievable doses.

  • Varian has RAPIDPLAN as the implementation of the system

  • Principle : Extract knowledge from a library of plans to generate a model that can predict the dose volume histogram.

  • Model originally developed by the Duke’s University.

Workflow of RAPIDPLAN KBP

Varian RAPIDPLAN : Model Building

  • Step 1: Define the case scenario where the RAPIDPLAN is to be used.

  • Step 2: Identify a set of cases satisfying the planning scenario : Uniform planning technique, and constraints.

  • Step 3: Create a RAPIDPLAN training dataset (note that the training dataset need not have plans that are created in VARIAN)

  • Step 4: Model training

  • Step 5: Model verification

  • Step 6: Model approval and deployment

Model training: Partitioning

  • Each OAR is partitioned into three parts. For each partition the following are calculated:

    • Volume of the partition

    • DVH of the partition

    • Geometry based expected dose (GED) : Dose to the voxel based on the pateint anatomy, dose to target and field position using distance from the voxel to the target surface.

Volume partitioning

Model Training: GED

  • Dose to voxels closer to the beam > dose to voxels away from the beam (inverse square law)

  • Dose to voxels where the beam fanline does not intersect the target is 0.

  • Dose to each voxel is obtained from the different beam orientations and summed to generate a GED distribution and GED DVH

Model Training : PCA

  • Principal component analysis (PCA) is a method to reduce the number of dimensions in the data.

  • Dimensions = variables (for DVH consider each volume bin to be a variable).

  • Doing a PCA is effectively reducing the number of volume bins to be considered for the future model to a single or few derived values.

  • Each in field volume partition –> obtain PC for DVH and GED

  • Take the first 2-3 PCS –> Regression (step wise)

  • Coefficients from the regression used to calculate the PC for the DVH

Example Gyn Planning Model

DVH Plot in RapidPlan

The extent of variation in the rectal doses can be seen in the above image.

Example Gyn Planning Model ..

RapidPlan Box Plots

The boxplots show the geometrical information in the OAR.

Example Gyn Planning Model ..

RapidPlan Residual Plot

This residual plot shows the distribution of the geometric information and the principal component of the DVH. Outlier cases can be detected using this plot

Is this sufficiently good ?

  • Yaorong Ge et al 4, reviewed the results reported by KBP studies.

  • Head neck IMRT cases 10 studies.

  • All studies demonstrated reduced OAR doses :

    • Dmax in serial organs reduced by about 6 - 8 Gy

    • Dmean in parallel organs reduced by about 4 - 5 Gy

  • Error in predicting optimal doses was low.

Visualization of KBP for Prostate Cancer

Deep learning

  • Possible to use deep learning to “learn” dose patterns from old plans

  • Use this to predict dose to different organs

  • Extension of the KBP approach

  • Several studies have demonstrated feasibility

Comparision of DL based dose prediction vs actual dose (Gronberg et al5)

Other niche tasks

  • Generate synthetic images (CT -> MRI or MRI -> CT) 6
  • Automated plan quality assurance 7
  • Deformable Image registration 8
  • Predict the need for online adaptive radiotherapy without doing segmenation9

Conclusion

  • KBP and AI in radiotherapy planning increasingly “practical”

  • Existing commercial KBP solutions however increase plan quality marginally

  • May be useful to predict the dose and do QA on “novice” plans

  • May be useful to define “achievable envelopes” of doses

  • May be useful in settings of online adaptive planning

  • Unlikely to outperform your experienced senior medical physicist any time soon.

Footnotes

  1. Moore KL. Automated radiotherapy treatment planning. Semin Radiat Oncol 2019;29:209–18. https://doi.org/10.1016/j.semradonc.2019.02.003.

  2. Ahmed S, Nelms B, Gintz D, Caudell J, Zhang G, Moros EG, et al. A method for a priori estimation of best feasible DVH for organs-at-risk: Validation for head and neck VMAT planning. Med Phys 2017;44:5486–97. https://doi.org/10.1002/mp.12500.

  3. Saha S, Sriram Prasath S, Arun B, Kalita SJ, Elavarasan N, Guha Adhya D, et al. ICON-P - A double-blind evaluation of quality improvements with individualized CONstraints from low-cost knowledge-based radiation therapy planning in prostate cancer. Tech Innov Patient Support Radiat Oncol 2023;26:100206. https://doi.org/10.1016/j.tipsro.2023.100206.

  4. Ge, Yaorong, and Q. Jackie Wu. 2019. “Knowledge-Based Planning for Intensity-Modulated Radiation Therapy: A Review of Data-Driven Approaches.” Medical Physics 46 (6): 2760–75. https://doi.org/10.1002/mp.13526.

  5. Gronberg MP, Gay SS, Netherton TJ, Rhee DJ, Court LE, Cardenas CE. Technical Note: Dose prediction for head and neck radiotherapy using a three-dimensional dense dilated U-net architecture. Med Phys 2021;48:5567–73. https://doi.org/10.1002/mp.14827.

  6. Boulanger M, Nunes J-C, Chourak H, Largent A, Tahri S, Acosta O, et al. Deep learning methods to generate synthetic CT from MRI in radiotherapy: A literature review. Phys Med 2021;89:265–81. https://doi.org/10.1016/j.ejmp.2021.07.027.

  7. Kalendralis P, Luk SMH, Canters R, Eyssen D, Vaniqui A, Wolfs C, et al. Automatic quality assurance of radiotherapy treatment plans using Bayesian networks: A multi-institutional study. Front Oncol 2023;13:1099994. https://doi.org/10.3389/fonc.2023.1099994.

  8. Zou J, Gao B, Song Y, Qin J. A review of deep learning-based deformable medical image registration. Front Oncol 2022;12:1047215. https://doi.org/10.3389/fonc.2022.1047215.

  9. Parchur AK, Lim S, Nasief HG, Omari EA, Zhang Y, Paulson ES, et al. Auto-detection of necessity for MRI-guided online adaptive replanning using a machine learning classifier. Medical Physics 2023;50:440–8. https://doi.org/10.1002/mp.16047.