Mach. Learn.: Sci. Technol. 10.1088/2632-2153/abb6d5)” width=”635″ height=”391″>

Successful radiation therapy relies on the creation of an accurate treatment plan that will deliver radiation dose precisely to the prescribed targets. The accuracy of this plan, however, is only as good as the accuracy of the underlying dose calculations. And for proton therapy, accurate dose calculation is even more critical, as protons deliver a more conformal dose distribution than photons and are more sensitive to anatomical changes.

Speaking at the Mayo Clinic’s 1st Proton Therapy Research Workshop, Steve Jiang – professor and director of the Medical Artificial Intelligence and Automation (MAIA) Laboratory at UT Southwestern Medical Center – described the key requirements of proton dose calculation – and described ways in which deep learning could help achieve these goals.

As well as high accuracy, Jiang explained, proton dose calculations also need to be fast. For treatment planning this means a few minutes; for replanning prior to fraction delivery in adaptive radiotherapy, a few seconds. Looking further ahead, we may see the introduction of real-time adaptation during treatment delivery. “We don’t do this right now,” he noted. “But at some point we may want to adapt the treatment plan in real time. For that kind of application, we will need dose calculation in milliseconds.”

Currently, there are two main types of technique used for dose calculation, represented by: pencil beam algorithms, which are less accurate but quite fast; and Monte Carlo (MC) simulations, which are more accurate but typically far slower. “But we need accuracy and speed for proton dose calculations,” said Jiang. “So there’s an unmet clinical need: we need to develop an algorithm that is both fast and accurate.”

So how can this be achieved? One approach is to improve the efficiency of MC calculations, using graphics processing units (GPUs) to accelerate MC code, for example, or deep learning-based denoising to reduce the noise inherent in MC-calculated results. Another option is to employ deep learning methods to improve the accuracy of pencil beam algorithms. Finally, it may be possible to develop new, totally different algorithms that meet both requirements; and deep learning could help explore this possibility.

Combining speed and accuracy

GPU-acceleration of MC simulations is already possible. Ten years ago (while at UC San Diego and in collaboration with Mass General Hospital), Jiang and colleagues developed gPMC, a MC package for fast proton dose calculation on a GPU. This enabled calculation of a typical proton treatment plan with 1% uncertainty in 10–20 s. Jiang notes that with today’s faster GPUs, gPMC may offer even higher efficiency.

Working with colleagues at the MAIA Lab, Jiang has also developed a deep learning-based MC denoiser. They created a deep dose plugin that can be added to any GPU-based MC dose engine to enable real-time MC dose calculation. The denoiser runs in just 39 ms, with the entire dose calculation taking just 150 ms. Jiang notes that the plugin was developed for photon beam radiotherapy, but could also be used for MC denoising in proton dose calculations.

Next, Jiang described ways to use deep learning techniques directly for dose calculation. He emphasized that this differs from dose prediction, which assumes a relationship between a patient’s anatomy and their optimal dose distribution, and uses this relationship to build a predictive model. After training on data from historical treatments of the same disease site, the model predicts an optimal dose distribution for the new patient and uses this to guide treatment planning. UT Southwestern has employed this type of patient-specific dose prediction clinically for over two years now.

But dose calculation is more than this. “Here, the relationship we are trying to exploit is between patient anatomy plus machine parameters and the actual dose distribution,” said Jiang. “You know the patient anatomy, you know the treatment plan, now you want to see what is the dose distribution, so it’s a dose calculation.”

Jiang’s team first developed the deep learning-based dose calculation model for photon beam radiotherapy. The model is trained using MC-calculated dose distributions for various patient anatomies and machine parameters. For the model inputs, the team used the patient CT scan and the ray tracing dose distribution for each beam, with machine parameters encoded into the ray tracing. “This makes the whole deep learning process easier and is a good way to incorporate physics into the deep learning,” Jiang noted.

The researchers applied a similar approach for proton dose calculation, using a deep learning model to boost the accuracy of pencil beam dose calculation to that of MC simulations. They trained and tested the model using pencil beam dose distributions and data from the TOPAS MC platform, for 290 head-and-neck, liver, prostate and lung cancer cases. For each plan, they trained the model to predict the MC dose distribution from the pencil beam dose.

The approach achieved high levels of agreement between the converted and the MC dose. “Compared with pencil beam, we see a huge improvement in accuracy, and the efficiency is still very high,” said Jiang. The developed model can be added to the clinical workflow of proton treatment planning to improve dose calculation accuracy.

Jiang also highlighted similar research under way by other groups, including DiscoGAN from Wuhan University, DKFZ’s use of artificial neural networks for proton dose calculation and deep learning-based millisecond speed dose calculation algorithm developed at Delft University of Technology.

Keeping users reassured

While deep learning may appear the obvious way forward for proton dose calculation, Jiang noted that people still feel more comfortable using physics-based models such as pencil beam algorithms and MC simulations. “When the idea of deep learning for dose calculation first came out, people had concerns,” he explained. “Because it’s data driven, not physics-based, you do not know when it’s going to fail; there might be unpredictable catastrophic failures. And because it’s a black box there’s no transparency.”

AI framework overcomes segmentation challenges for online adaptive radiotherapy

The answer may lie in hybrid models, such as the examples described above that use pencil beam or ray tracing data as inputs to a deep learning model. Here, the physics (machine parameters) is encoded in the input data, which already has an accuracy of 80–90%. Deep learning can then address effects such as scatter and inhomogeneity to gain the remaining 20% accuracy that’s very difficult to achieve with analytical algorithms. This should provide both the desired accuracy and efficiency.

“I actually think this is a good idea because it can also eliminate unpredictable, catastrophic failures,” Jiang concluded. “I’d feel much more comfortable with the results. Also you’d have some degree of transparency, because you know the first order primary effect that is there is physics-based, and that’s correct.”