[Paper Review] An Artificial Intelligence-Driven Agent for Real-Time Head-and-Neck IMRT Plan Generation using Conditional Generative Adversarial Network (cGAN)

Xinyi Li, Yang Sheng+7 more|arXiv (Cornell University)|Sep 27, 2020

Advanced Radiotherapy Techniques34 references36 citations

TL;DR

This paper presents an AI-driven agent based on a conditional GAN (cGAN) that enables fully automated, real-time head-and-neck IMRT plan generation in under 3 seconds. Using a novel pyramid-structured generator (PyraNet) and a customized DenseNet discriminator, the system predicts 9 fluence maps simultaneously from 3D CT data, achieving dosimetric quality comparable to clinical TPS plans with no significant differences in critical structure doses.

ABSTRACT

Purpose: To develop an Artificial Intelligence (AI) agent for fully-automated rapid head and neck (H&N) IMRT plan generation without time-consuming inverse planning.$$$$ Methods: This AI agent was trained using a conditional Generative Adversarial Network architecture. The generator, PyraNet, is a novel Deep Learning network that implements 28 classic ResNet blocks in pyramid-like concatenations. The discriminator is a customized 4-layer DenseNet. The AI agent first generates customized 2D projections at 9 template beam angles from 3D CT volume and structures of a patient. These projections are then stacked as 4D inputs of PyraNet, from which 9 radiation fluence maps are generated simultaneously. Finally, the predicted fluence maps are imported into a commercial treatment planning system (TPS) for plan integrity checks. The AI agent was built and tested upon 231 oropharyngeal plans from a TPS plan library. Only the primary plans in the sequential boost regime were studied. A customized Harr wavelet loss was adopted for fluence map comparison. Isodose distributions in test AI plans and TPS plans were qualitatively evaluated. Key dosimetric metrics were statistically compared.$$$$ Results: All test AI plans were successfully generated. Isodose gradients outside of PTV in AI plans were comparable with TPS plans. After PTV coverage normalization, $D_{mean}$ of parotids and oral cavity in AI plans and TPS plans were comparable without statistical significance. AI plans achieved comparable $D_{max}$ at 0.01cc of brainstem and cord+5mm without clinically relevant differences, but body $D_{max}$ was higher than the TPS plan results. The AI agent needs ~3s per case to predict fluence maps.$$$$ Conclusions: The developed AI agent can generate H&N IMRT plans with satisfying dosimetry quality. With rapid and fully automated implementation, it holds great potential for clinical applications.

Motivation & Objective

• To eliminate time-consuming inverse planning in head-and-neck IMRT by developing a fully automated AI-driven planning solution.
• To address the clinical need for rapid, consistent, and high-quality IMRT plan generation in the sequential boost regimen for oropharyngeal cancer.
• To achieve real-time plan generation without compromising target coverage or organs-at-risk sparing.
• To validate the AI-generated plans against clinical TPS plans using quantitative and qualitative dosimetric metrics.

Proposed method

• A conditional Generative Adversarial Network (cGAN) architecture was employed, with a novel generator network (PyraNet) based on 28 ResNet blocks arranged in pyramid-like concatenations.
• The discriminator is a 4-layer customized DenseNet designed to distinguish real from generated fluence maps.
• The AI agent first creates 2D beam projections at 9 standard angles from a 3D CT volume and structure set.
• These 2D projections are stacked into a 4D input tensor fed into PyraNet to generate 9 simultaneous fluence maps.
• A customized Harr wavelet loss function was used to improve fluence map similarity between AI and TPS plans.
• Generated fluence maps were imported into a commercial TPS for integrity checks and dosimetric evaluation.

Experimental results

Research questions

• RQ1 Can an AI agent generate head-and-neck IMRT plans with dosimetric quality matching clinical TPS plans in real time?
• RQ2 How does the proposed cGAN-based method perform in maintaining target coverage and organs-at-risk sparing compared to clinical plans?
• RQ3 What is the inference speed of the AI agent, and can it support clinical workflow integration?
• RQ4 To what extent do the AI-generated fluence maps preserve isodose gradient characteristics compared to TPS plans?

Key findings

• All 231 test cases were successfully generated within approximately 3 seconds per case, demonstrating real-time feasibility.
• Isodose gradients outside the PTV in AI-generated plans were qualitatively comparable to those in clinical TPS plans.
• After PTV coverage normalization, mean doses to parotids and oral cavity showed no statistically significant difference between AI and TPS plans.
• Maximum doses at 0.01 cc for brainstem and cord+5 mm were comparable between AI and TPS plans, with no clinically relevant differences.
• Body maximum dose was slightly higher in AI plans than in TPS plans, though not clinically concerning.
• The use of a customized Harr wavelet loss improved fluence map fidelity, contributing to consistent dosimetric outcomes.