A Call for a Collaborative Framework for Automation in Adaptive Radiotherapy

Abstract

Adaptive radiotherapy (ART) has been shown to improve geometric and dosimetric accuracy, with emerging evidence for clinical benefit, but it remains resource-intensive and lacks scalability. This limitation arises from multiple factors, including the complexity of current systems, the closed and proprietary nature of radiotherapy platforms, and the need for human oversight driven in part by clinical risk considerations. Historically, major advances in radiotherapy—from Intensity-Modulated Radiation Therapy (IMRT) and Image-Guided Radiation Therapy (IGRT) to Magnetic Resonance-guided Radiotherapy (MRgRT) and Deep Learning in Radiotherapy (DLinRT) (particularly for auto-contouring)—have thrived through open collaboration and transparency. The community can accelerate ART innovation by returning to this model. Open-source initiatives such as Computational Environment for Radiotherapy Research (CERR), Open Knowledge-based Planning (OpenKBP), and matRad demonstrate how shared tools and methods improve reproducibility and drive scientific progress. The next critical step is to develop collaborative, structured frameworks that enable safe, secure interaction between academic and vendor systems—safeguarding intellectual property while fostering co-development and validation. Through structured transparency and shared accountability, the radiotherapy field can transform automation from a closed, non-transparent architecture into a collective learning ecosystem, ultimately extending the life-saving benefits of ART to more patients worldwide through openness, trust, and collective innovation.

Introduction: Adaptive Radiotherapy as Clinical Reality

Online adaptive radiotherapy (ART) ¹ has moved beyond proof of concept—it is now a clinically implemented workflow with demonstrated dosimetric and workflow benefits, and emerging clinical evidence of therapeutic value in selected settings. As summarized in a recent review paper, ² evidence from recent clinical experience indicates that online ART can enhance both target coverage and organ-at-risk sparing across conventional and stereotactic body radiotherapy, with particularly notable benefits reported in head and neck, abdominal/pelvic, and ultracentral lung treatments. MRI-guided systems such as the Elekta Unity MR-Linac (Stockholm, Sweden) and Cone-Beam CT (CBCT)-based systems such as Varian Ethos (Palo Alto, CA, USA) now demonstrate that high-quality adaptive planning can be performed directly on the treatment couch—marking a major step toward truly personalized radiotherapy.

Despite this clinical progress, the workflow remains resource intensive. ² A typical adaptive session requires multiple professionals: physicians manually review and modify propagated contours; dosimetrists adjust the plan to meet clinical objectives; and physicists perform on-the-fly quality assurance before delivery. Each step adds time and complexity, limiting scalability—especially in centers with constrained staffing or experience.

Thus, the field faces a paradox: the technology for online adaptation exists, but the human workflow constrains its reach. For adaptive radiotherapy to become standard practice rather than a specialized capability, the community must accelerate automation across contouring, planning, and plan verification. A schematic overview of the current ART workflow and its key bottlenecks is shown in Figure 1.

Figure 1.

Current adaptive radiotherapy (ART) workflow and key bottlenecks (dashed line, orange color)

This editorial article aims to spark discussion on how vendors, clinicians, and researchers can collaborate to develop robust, transparent automation—enabling faster, more accessible adaptive workflows that maintain safety and quality while reducing manual intervention. By rethinking how algorithms are designed and shared, the field can extend the benefits of adaptation to more patients and institutions worldwide.

A Historical Continuum of Innovation

The path to adaptive radiotherapy has been shaped by a continuous progression of innovations, with key technological milestones illustrated in Figure 2.

Figure 2.

Major technological breakthroughs in the field of radiation therapy (The years shown correspond to the publication dates of the cited papers, which may not reflect the exact time of first invention or initial clinical implementation.)

Intensity-Modulated Radiation Therapy (IMRT) ³ introduced inverse planning, shifting the field from manual beam adjustments to algorithm-driven optimization. For the first time, planners could encode clinical goals and let computation determine optimal solutions. Techniques such as IMRT, Volumetric Modulated Arc Therapy (VMAT), ⁴ and AI-based methods are driven by advances in computational algorithms, while their clinical implementation relies critically on innovations in delivery hardware, imaging systems, and integrated treatment platforms. VMAT further extended this paradigm by demonstrating that complex, high-quality plans could be optimized and delivered efficiently within minutes. Image-Guided Radiation Therapy (IGRT) emerged with the clinical adoption of CBCT, ⁵ bringing daily volumetric imaging into the treatment room and enabling real-time anatomical verification before each fraction.

The rise of deep learning marked a new wave of progress. Early convolutional architectures such as U-Net, ⁶ followed by transformer-based models, formed the foundation of Deep Learning in Radiotherapy (DLinRT). These models revolutionized segmentation, ⁷ contouring, ⁸ and image-based decision making, reducing hours of expert effort to near-instant automation—laying essential groundwork for on-couch ART workflows.

In parallel, Magnetic Resonance-guided Radiotherapy (MRgRT) ⁹ advanced IGRT imaging capabilities by introducing superior soft-tissue contrast and real-time visualization. MRI-guided linacs enabled precise tracking of anatomical motion and facilitated the first clinically feasible implementations of daily online adaptation.

Each of these milestones emerged from the close interaction of academic research, industrial development, and clinical implementation, reflecting a continuous co-evolution across these domains.

Lessons From AI: The Power of Open Ecosystems

The most rapid technological advances of the past two decades have stemmed not from hardware alone but from a culture of openness. CUDA ¹⁰ transformed GPUs from gaming devices into engines of scientific computing. On this foundation, open frameworks such as TensorFlow ¹¹ and PyTorch ¹² created shared infrastructures where models, code, and data circulated freely—turning AI into a global collaborative enterprise.

Such transparency has helped facilitate knowledge sharing, benchmarking, and cross-institutional validation, contributing to the pace of innovation. Architectures like U-Net ⁶ and its transformer-based successors achieved expert-level segmentation accuracy, while shared datasets and benchmarks have improved opportunities for comparison and validation, reproducibility in complex machine learning systems remains an ongoing challenge.

The same principle applies to radiotherapy: when algorithms, data, and interfaces are designed for structured transparency, innovation compounds across institutions. Open APIs, standardized validation datasets, and modular sandboxes can stimulate the collaborative momentum that transformed AI—when designed within appropriate regulatory, safety, and intellectual property frameworks. This echoes the principle of open innovation articulated by Chesbrough ¹³ : valuable knowledge is distributed widely, and advancement depends on mechanisms that allow it to flow freely. The considerations of regulatory, safety and intellectual property are discussed in a subsequent section.

Hope and Limitations in the Current Landscape

Automation in radiotherapy is advancing rapidly. Proprietary systems such as Varian’s Intelligent Optimization Engine (IOE) and Elekta’s Warm Start Optimization (WSO) have demonstrated that high-quality adaptive plans can be generated within minutes, translating complex optimization into practical, reproducible clinical tools. These systems represent genuine achievements: automation that works at scale, delivering consistency and speed beyond what manual workflows can achieve.

Yet, their success also highlights the field’s central limitation—opacity. The algorithms that drive these systems are embedded within closed architectures. Their internal implementation details are typically not accessible for external evaluation or modification, which can limit opportunities for independent benchmarking and systematic comparison. As a result, validation and refinement often rely on empirical testing and performance evaluation, rather than structured, standardized benchmarking across systems. Importantly, access to algorithms does not necessarily confer full interpretability, particularly in complex machine learning models. The AAPM Task Group 332 report ¹⁴ provides valuable methods for verifying such “black boxes,” but coping with opacity is not a sustainable path forward. For automation to mature, the field must shift from validation around the algorithm to collaboration within it.

Encouragingly, there are glimpses of a more open future. Academic initiatives like GPU based Multi-Criteria Optimization (gMCO) ¹⁵ and Expedited Constrained Hierarchical Optimization (ECHO) ¹⁶ have shown how transparency and shared methods can accelerate innovation. gMCO, built on open GPU frameworks such as CUDA and informed by early GPU research from Wetzl et al, ¹⁷ generates thousands of Pareto-optimal plans in seconds. ECHO applies sequential convex programming to clinical IMRT/VMAT planning and releases its code publicly, enabling reproducibility and independent evaluation. Projects such as Computational Environment for Radiotherapy Research (CERR), ¹⁸ Open Knowledge-based Planning (OpenKBP), ¹⁹ and matRad ²⁰ further promote open datasets, dose engines, and benchmarking—laying the groundwork for standardized validation across institutions. To be specific, CERR ¹⁸ provides a unified platform for importing, analyzing, and comparing radiotherapy plans—supporting reproducible algorithm development across institutions. OpenKBP ¹⁹ supplies standardized datasets, objectives, and benchmarks for knowledge-based planning, fostering reproducible model evaluation and cross-center comparison. MatRad ²⁰ offers an open, MATLAB-based planning system with transparent dose engines and optimization models, enabling rapid prototyping of new planning strategies.

Extending this collaborative model to treatment planning—through shared APIs, standardized datasets, and modular optimization frameworks—can similarly accelerate innovation, while maintaining patient safety through appropriate validation, regulatory compliance, and clinical oversight. Commercial platforms also show promise. Eclipse’s Scripting API (ESAPI) and RayStation’s Python interface allow limited programmatic access to planning data and workflow automation, enabling physicists to support innovation within controlled research and development environments, subject to appropriate validation and governance. Such interfaces are primarily intended for research and controlled integration and do not, in themselves, constitute clinically approved tools. These are important steps toward structured transparency, but their reach remains limited. The optimization logic and dose engines remain closed, preventing researchers from probing or refining the decision layers that govern plan quality.

Regulatory, Safety, and Economic Constraints

While the limitations of current systems highlight the need for greater collaboration, it is equally important to recognize the structural constraints that shape their design and adoption.

First, radiotherapy technologies operate within stringent regulatory frameworks (e.g., FDA clearance and CE marking), where even incremental modifications to planning algorithms or dose engines may require substantial re-validation. For vendors, this creates a high barrier to exposing internal components, as increased openness may introduce regulatory complexity and risk. Any proposed collaborative framework must therefore be compatible with existing regulatory pathways, rather than assuming unrestricted access to clinical systems.

Second, the persistence of human-intensive workflows may reflect not only technical limitations but also clinical risk aversion. Even when automated planning systems demonstrate strong dosimetric performance, clinicians often prefer to retain manual oversight, particularly in high-stakes treatment decisions. This highlights that trust, interpretability, and validation are as critical as algorithmic performance in enabling broader adoption of automation.

Finally, economic incentives play a central role in shaping the feasibility of collaboration. Academic research is typically driven by publication and grant funding, whereas industrial innovation is closely tied to product development, market differentiation, and financial sustainability. From this perspective, increased openness must be aligned with clear commercial value.

Rather than requiring full transparency, collaborative approaches can be structured to provide mutual benefit—for example, by enabling standardized benchmarking that reduces internal validation burden, facilitating faster integration of externally developed methods, and supporting regulatory-aligned interfaces that lower barriers to product development and deployment.

Emerging efforts in medical imaging and radiotherapy further suggest that such collaboration can be supported through hybrid funding models. Public research funding agencies (e.g., NIH/NCI, MRC, CIHR, INSERM, AMED) provide support for foundational research and shared data infrastructure, while cloud-based platforms enable scalable access to data, models, and computational resources. In these settings, industry participation is often driven not by unrestricted openness, but by opportunities to provide infrastructure, services, and deployment pathways. This model illustrates that collaboration does not require the absence of profit motives; rather, it can be sustained by aligning incentives across stakeholders, where shared infrastructure reduces duplication of effort, accelerates innovation, and preserves commercial viability.

Co-Creating the Future: A Collaborative Roadmap

To accelerate progress, the field must bridge the gap between promise and practical limitations. Instead of full open-source exposure, the future lies in collaborative openness: modular frameworks that allow academic algorithms to be evaluated, benchmarked, and refined within vendor systems. This approach protects intellectual property while enabling scientific validation, strengthening both research and clinical reliability. As automated adaptive planning grows, the community needs transparency, extensibility, and verifiability to ensure innovation is shared rather than siloed. A practical next step is for stakeholders across disciplines to establish a standardized, sandboxed API layer—developed through coordinated, multi-stakeholder collaboration involving academia, industry, and clinical organizations with shared governance structures and clearly defined roles for each stakeholder group—to support reproducible benchmarking, safe plugin development, and cross-platform validation. Existing collaborative initiatives provide useful insights into structured data sharing and governance, while also highlighting the practical challenges of access, transparency, and stakeholder alignment. Such an interface could enable curated datasets and simplified computational tools to be shared within a secure, non-punitive environment, allowing the community to iterate and improve algorithms while preserving the safety and integrity of proprietary clinical systems.

Conclusion: From Closed to Collaborative

Automation is already reshaping radiotherapy. The task now is to ensure automated adaptive workflows evolve with responsibility, transparency, and inclusivity. Proprietary systems have proven feasibility; the future depends on making automation verifiable and continually improvable. Progress will come from collaborative openness, where the community co-develops standards for contouring, optimization, and plan verification—shifting from isolated, opaque workflows toward more structured, evaluable, and collaboratively improved systems through iterative, evidence-based improvements rather than complete system transparency. Adaptive radiotherapy’s future will be defined by those who open algorithms wisely and advance patient care through shared intelligence and trust.