[Paper Review] Automated Treatment Planning in Radiation Therapy using Generative Adversarial Networks

Rafid Mahmood, Aaron Babier+3 more|arXiv (Cornell University)|Jul 17, 2018

Radiomics and Machine Learning in Medical Imaging69 citations

TL;DR

The paper introduces a GAN-based knowledge-based planning pipeline that predicts 3D dose distributions directly from contoured CT slices to automate radiotherapy treatment planning for oropharyngeal cancer, followed by an optimization stage to produce deliverable plans.

ABSTRACT

Knowledge-based planning (KBP) is an automated approach to radiation therapy treatment planning that involves predicting desirable treatment plans before they are then corrected to deliverable ones. We propose a generative adversarial network (GAN) approach for predicting desirable 3D dose distributions that eschews the previous paradigms of site-specific feature engineering and predicting low-dimensional representations of the plan. Experiments on a dataset of oropharyngeal cancer patients show that our approach significantly outperforms previous methods on several clinical satisfaction criteria and similarity metrics.

Motivation & Objective

• Motivate automation of radiation therapy treatment planning to reduce manual effort and increase consistency.
• Develop a knowledge-based planning pipeline that learns dose distributions directly from CT data without heavy feature engineering.
• Demonstrate that a GAN-based approach can generate deliverable plans that meet clinical criteria across OARs and targets.

Proposed method

• Recast KBP dose prediction as an image colorization problem using a pix2pix-style GAN with a U-net generator.
• Train on contoured CT slices from 130 patients to predict 2D dose slices, assembling a 3D dose distribution.
• Use a discriminative network to critique generated dose distributions, mimicking oncologist evaluation.
• Pass GAN predictions into an inverse optimization pipeline to produce deliverable plans that satisfy clinical constraints.
• Optimize with a forward model minimizing 65 objective functions across OARs and targets using Gurobi.
• Evaluate against baselines (BQ, gPCA, RF, CNN) using clinical criteria and gamma passing rate.

Experimental results

Research questions

• RQ1 Can a GAN learn to predict clinically acceptable 3D dose distributions directly from CT anatomy without engineered features?
• RQ2 Do GAN-generated dose predictions, when optimized, produce deliverable treatment plans that satisfy clinical criteria as well as or better than existing KBP methods?
• RQ3 How does GAN-based KBP compare to traditional feature-based and voxel-based approaches in plan quality and deliverability?
• RQ4 Is the GAN-based KBP pipeline generalizable to other cancer sites beyond oropharyngeal cancer?
• RQ5 What is the impact of incorporating an inverse optimization stage on plan deliverability and clinical criterion satisfaction?

Key findings

• GAN plans outperformed baseline KBP methods (BQ, gPCA, RF, CNN) on clinical criteria satisfaction.
• GAN plans achieved the best overall balance between OAR sparing and target coverage among evaluated methods.
• Gamma passing rate analyses show GAN plans most closely resemble clinical dose distributions on average, especially around targets.
• GAN plans sometimes delivered lower OAR doses than clinical plans, improving OAR criteria satisfaction.
• CNN was the closest competitor to GAN, with GAN maintaining a small but consistent advantage.