DAART: a deep learning platform for deeply accelerated adaptive radiation therapy for lung cancer

Hamed Hooshangnejad, Quan Chen, Xue Feng, Rui Zhang, Reza Farjam, Khinh Ranh Voong, Russell K. Hales, Yong Du, Xun Jia, Kai DingFrontiers in Oncology 2023 · DOI

TL;DR

The DAART platform uses a 3D generative adversarial network (GAN) to synthesize radiotherapy planning images directly from initial diagnostic scans. By eliminating the need for a separate "simulation" imaging session, this approach can reduce the wait time between lung cancer diagnosis and treatment initiation by 50% (from four weeks down to two).

What Problem Does This Paper Address?

In the world of oncology, time is tissue. For patients with non-small cell lung cancer (NSCLC), the current workflow for radiation therapy (RT) is notoriously slow, with a median delay of four weeks from diagnosis to the first beam of radiation.

The bottleneck isn't just scheduling; it’s physics. Diagnostic scans (PET/CT) are usually taken on curved couches with the patient breathing naturally. However, for precise radiation delivery, doctors need a "Planning CT" (pCT) taken on a flat couch, often using specific motion-management techniques like breath-holding. This "simulation" step requires a separate appointment, specialized staffing, and complex manual processing. During this month-long gap, tumors can grow or shift, potentially requiring a complete restaging of the cancer. The authors developed DAART (Deeply Accelerated Adaptive Radiation Therapy) to bridge this gap using deep learning.

Methods

The core of the DAART platform is a system called deepPERFECT. It utilizes a 3D Pix2Pix Generative Adversarial Network (GAN) to transform a diagnostic PET/CT into a synthesized planning CT (sCT).

Key technical innovations include:

  • Predicting Motion, Not Pixels: Instead of having the GAN "hallucinate" new CT intensities, the model generates Deformation Vector Fields (DVFs). It essentially learns how to "warp" the diagnostic scan into the treatment position. This preserves the original Hounsfield Units (HU), which are critical for accurate radiation dose calculations.
  • Contrast Fidelity Loss: The authors introduced a novel loss function that mimics how physicians view scans. By applying a differentiable "window/level" function during training, the AI is forced to prioritize the similarity of features that are clinically relevant, such as pulmonary vessels and lung boundaries.
  • Diffeomorphic Transformation: The system ensures the warping is "physically plausible"—meaning it prevents the AI from "tearing" or "folding" the digital anatomy, maintaining the topological integrity of the patient's organs.

Datasets

The study used retrospective data from 15 patients with early-stage NSCLC who had already undergone standard-of-care SBRT (Stereotactic Body Radiation Therapy).

  • Training: 10 cases were used to train the GAN.
  • Testing/Virtual Trial: 5 cases were reserved to evaluate the model's performance and conduct a "virtual clinical trial" to see if the AI-generated plans were actually safe for treatment.
  • Preprocessing: All scans were resampled to a 2.5 mm voxel dimension to accommodate GPU memory limitations.

Key Insights and Ideas

  • The "Aha" Moment on DVFs: Most image-to-image translation tasks in AI focus on pixel-wise similarity. However, in medical physics, the intensity of a pixel (the HU value) represents physical density. By predicting the deformation rather than the intensity, the researchers ensured that the physics of the dose calculation remained grounded in the original diagnostic data.
  • Clinical Realism over Mathematical Perfection: Standard L1 or L2 loss functions often result in blurry images. The "Contrast Fidelity Loss" is a clever way to encode clinical expertise into the neural network, telling the model: "It matters more if you get the lung-vessel boundary right than if you get the background noise perfect."
  • Adaptive Safety Net: The authors don't suggest treating blindly. The sCT is used to create a "preliminary plan." On the first day of treatment, the plan is refined using "Adapt to Position" (ATP) or "Adapt to Shape" (ATS) methods, ensuring the speed of AI is balanced with the precision of real-time imaging.

Key Results

  • High Fidelity: The synthesized CTs were remarkably accurate, with a mean intensity fidelity of 14 HU for the lungs (a very low error margin in CT imaging).
  • Geometric Accuracy: The Dice Similarity Coefficient (DSC)—a measure of overlap—was 0.97 for the lungs and 0.91 for the body, indicating the AI successfully "warped" the patient into the correct treatment posture.
  • Time Savings: The workflow reduces the time-to-treatment from 4 weeks to 2 weeks, a 50% reduction that significantly lowers the risk of tumor progression during the wait.
  • Plan Quality: In the virtual clinical trial, the plans designed on AI-synthesized scans were "non-inferior" to standard plans, meeting all clinical safety constraints for protecting healthy organs like the heart and spinal cord.

Limitations and Future Work

The primary limitation is the sample size; 15 patients is a small cohort for a deep learning study, though the results were statistically significant for this proof-of-concept. Additionally, due to GPU hardware constraints, the images had to be down-sampled, which might obscure very fine anatomical details.

The authors suggest that future work will involve multi-institutional trials with larger datasets and the use of full-resolution CT scans. They also aim to extend the deepPERFECT model to other anatomical sites beyond the lungs, such as the head, neck, and brain.

Summary updated

Abstract

DAART reduces the common online ART (ART) treatment course by at least two weeks, resulting in a 50% shorter time to treatment to lower the chance of restaging and loss of local control.
BibTeX
@inproceedings{hooshangnejad2023daart,
  author = {Hamed Hooshangnejad and Quan Chen and Xue Feng and Rui Zhang and Reza Farjam and Khinh Ranh Voong and Russell K. Hales and Yong Du and Xun Jia and Kai Ding},
  title = {{DAART: a deep learning platform for deeply accelerated adaptive radiation therapy for lung cancer}},
  booktitle = {Frontiers in Oncology},
  year = {2023},
  doi = {10.3389/fonc.2023.1201679},
}
Figure