Accelerating Dose Calculation and Planning for Adaptive Radiotherapy

26 concepts19 papers1 proposal

During a course of radiation therapy, a patient’s internal anatomy is constantly in flux. Tumors shrink, organs shift due to digestion, and breathing causes continuous cyclical movement. To maximize tumor coverage and avoid damaging surrounding healthy tissue, clinicians ideally need online adaptive radiotherapy (ART)—the ability to recalculate and adjust the radiation beam parameters while the patient is lying on the treatment table.

However, traditional radiotherapy planning workflows are plagued by severe computational latency. The processes of segmenting organs from daily images, simulating radiation transport, and optimizing beam intensities can take anywhere from twenty minutes to several hours. This latency forces a compromise: clinicians must use large safety margins around the tumor, inadvertently exposing healthy tissue to radiation.

To achieve true real-time adaptation, the entire planning pipeline must be compressed into a window of less than three minutes. A rapidly growing body of research addresses this bottleneck by replacing slow, iterative physics simulations and manual workflows with ultra-fast deep learning surrogates, hardware acceleration, and physics-informed optimization.


Speeding Up the Gold Standard: Dose Calculation

To evaluate whether a radiotherapy plan is safe, systems must calculate the "dose distribution"—how much energy the radiation beams will deposit across different tissues. The gold standard for this calculation is Monte Carlo (MC) simulation, which models the stochastic transport of millions of individual particles through matter.

Because tracking millions of particle histories is computationally expensive, standard CPU-based implementations can take hours, as noted by M. Reza Bakhtiari et al.. While leveraging the parallel processing power of GPUs provides substantial acceleration, hardware optimization alone cannot easily bridge the gap to millisecond-level real-time constraints.

To bypass the brute-force tracking of particles, researchers have introduced mathematical and physical shortcuts:

  • Analytical Approximations: Instead of simulating individual particles, Mark Bangert et al. introduced Analytical Probabilistic Modeling (APM). By approximating proton dose curves as a sum of Gaussians, the authors turned complex uncertainty analysis into a closed-form equation, allowing clinicians to calculate dose variance without running thousands of distinct simulations.
  • Physics-Inspired Exploitation of Locality: Radiation particles have a finite physical range; they lose energy and stop after traveling a certain distance. Francesca De Benetti et al. exploited this property in their DosePatch framework. Instead of performing a whole-body simulation, the authors cropped medical images into overlapping patches scaled to the physical range of the particles. Because distant voxels do not physically interact, this localized approach yields a 10x speedup while maintaining high accuracy.

Deep Learning as a Physics Surrogate

The most dramatic acceleration comes from shifting the heavy computational burden from online inference to offline training. By training deep neural networks on thousands of pre-calculated plans, these models learn to predict dose distributions almost instantaneously.

Instead of treating dose prediction as a generic image-to-image translation task, recent architectures explicitly model the underlying physics of particle transport. For example, Oscar Pastor-Serrano et al. framed particle transport as a sequence-modeling task. Their Dose Transformer Algorithm (DoTA) treats the penetration of a radiation beam through successive tissue layers similarly to words in a sentence, achieving Monte Carlo-level accuracy while running 4,000 times faster.

Traditional Monte Carlo:
[Particle Source] ──> (Track Millions of Stochastic Trajectories) ──> [Dose Map] (Hours)

Deep Learning Surrogate:
[Patient Anatomy + Beam Angles] ──> (Transformer / GAN Inference) ──> [Dose Map] (Milliseconds)

Similarly, Lian Zhang et al. combined "beam masks" with a sliding window training technique to capture both local and global particle interactions, predicting pencil beam scanning proton therapy doses in under 0.25 seconds.

A persistent challenge in training these surrogates is the scarcity of high-precision Monte Carlo data required for ground truth. To overcome this, Sonia Martinot et al. utilized a weakly supervised "noise-to-noise" training strategy with 3D ConvLSTMs. This allowed the network to learn high-precision dose distributions directly from computationally cheap, low-precision, noisy simulations, bypassing the need for pristine training datasets.


Bypassing Inverse Optimization

Once a system can calculate a dose, it must solve the inverse optimization problem: determining the exact machine parameters (such as beam weights, shapes, or catheter placements) required to deliver that desired dose. Traditional optimization algorithms are highly iterative and slow, particularly for complex modalities like proton arc therapy (Bas A. de Jong et al.) or high-dose-rate brachytherapy (Krzysztof L. Sadowski et al.).

To accelerate this step, researchers are pursuing two distinct paths: faster mathematical solvers and direct parameter prediction.

Faster Mathematical Solvers

By restructuring the underlying optimization mathematics, researchers can achieve massive speedups without altering the clinical goals. For rotating shield brachytherapy, Myung Cho et al. implemented a Proximal Graph Solver (POGS). This mathematical framework splits complex constraints into simpler sub-problems that can be solved in parallel, resulting in an 18-fold reduction in optimization time.

Direct Parameter Prediction via GANs

The more radical approach eliminates the iterative optimization loop entirely. Laura Buchanan et al. trained conditional Generative Adversarial Networks (cGANs) to map a target dose distribution directly to deliverable machine parameters. Instead of guessing parameters, calculating the dose, and adjusting iteratively, the generative network learns the direct inverse mapping, collapsing a 20-minute optimization process into less than 20 seconds.


End-to-End Automation and Clinical Integration

Achieving real-time adaptive radiotherapy requires automating every step of the workflow, ensuring that no manual bottlenecks remain.

[Daily Patient Scan] 
       │
       ▼
[Auto-Segmentation]  <── Deep Learning Contouring (Erdur et al.)
       │
       ▼
[Direct Auto-Planning] <── Generative Parameter Prediction (Buchanan et al.)
       │
       ▼
[Physics-Guided Check] <── Soft Constraints & Growth Models (Balcerak et al.)
       │
       ▼
[Immediate Delivery]

First, the patient's daily anatomy must be segmented. As reviewed by Ayhan Can Erdur et al., deep learning autosegmentation models can now delineate tumors and organs-at-risk in seconds, replacing hours of manual contouring by clinicians. Sijuan Huang et al. demonstrated an integrated system that pairs auto-segmentation directly with auto-planning, moving the field closer to a unified "one-shot" clinical workflow.

To streamline the workflow further, the DAART platform developed by Hamed Hooshangnejad et al. uses a 3D GAN to synthesize radiotherapy planning images directly from initial diagnostic scans. This eliminates the need for a dedicated, time-consuming "simulation" imaging session entirely, drastically reducing the time from diagnosis to first treatment.

Finally, because these ultra-fast models operate as black boxes, researchers embed physical and biological guardrails to ensure patient safety. Stefanos Achlatis et al. designed a two-stage deep learning approach that incorporates physics-guided dose predictions directly into the network's loss function. Taking a biological approach, Michal Balcerak et al. introduced the GliODIL framework, which treats biological tumor growth equations as flexible "soft constraints" within a machine learning solver. This ensures that the generated plans are not only computationally instantaneous but also strictly adhere to the physical and biological realities of cancer progression.

Key Foundational Work

Applying graphics processor units to Monte Carlo dose calculation in radiation therapy (2010) — contributed 3 concept elements to this direction.
Analytical probabilistic proton dose calculation and range uncertainties (2014) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_real_time_adaptive_radiotherapy_constrai_09c1996f,
  title = {real-time adaptive radiotherapy constraints},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/real-time-adaptive-radiotherapy-constraints/rdr_09c1996f5e7e4fb6a0ba73fd9354faee},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_real_time_adaptive_radiotherapy_constrai_09c1996f_v1,
  title = {real-time adaptive radiotherapy constraints},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/real-time-adaptive-radiotherapy-constraints/rdr_09c1996f5e7e4fb6a0ba73fd9354faee/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake