Radiation therapy has evolved significantly over recent decades, with technologies like intensity-modulated radiation therapy (IMRT), stereotactic radiosurgery, image-guided radiation therapy (IGRT), and adaptive radiation therapy (ART) improving precision in targeting tumors. These advances allow clinicians to deliver higher radiation doses to tumors while sparing nearby healthy tissue, but each technique comes with its own planning complexity and quality assurance demands.
Global healthcare pressures are driving interest in AI for radiation oncology. An aging population means more cancer cases, while healthcare systems face pressure to improve outcomes and control costs. AI offers a way to handle the growing patient load more efficiently by automating time-consuming planning tasks, reducing treatment preparation time, and supporting more personalized treatment decisions.
This 2024 narrative review from Asan Medical Center, South Korea, examines where AI stands in radiation oncology today, what opportunities it opens up, and what barriers still stand in the way of widespread clinical adoption. It covers four main application areas: treatment planning, image analysis, adaptive therapy, and predictive analytics.
One of the most immediate applications of AI in radiation oncology is automating treatment planning. Machine learning algorithms can analyze large datasets of past treatment plans and patient outcomes to suggest optimal radiation parameters, reducing the hours-long manual planning process to minutes while maintaining or improving plan quality.
A specific challenge in treatment planning is fluence map optimization, which determines the intensity and direction of radiation beams to maximize tumor dose and minimize exposure to surrounding healthy tissue. Deep learning models can now predict optimal fluence patterns based on each patient's anatomy and tumor characteristics, dramatically reducing the need for manual adjustment by dosimetrists.
Commercial systems like Varian's Ethos Therapy and Elekta's MOSAIQ Plaza already integrate AI for automated planning. These platforms use historical data and patient-specific inputs to generate treatment plans that can be reviewed and approved by radiation oncologists, with the human expert serving as a quality check rather than the primary planner.
Before radiation can be delivered, physicians must precisely define the tumor boundaries and identify nearby organs that need to be protected. This process, called segmentation, has traditionally been done by hand and takes considerable time while also varying between individual clinicians. AI-based segmentation tools offer a consistent, rapid alternative.
Deep learning models, particularly convolutional neural networks (CNNs) and U-Net architectures, have become the standard approach for automated medical image segmentation. These models are trained on thousands of annotated scans from CT, MRI, and PET imaging and can accurately delineate tumor boundaries and surrounding organs within seconds, closely matching the accuracy of expert radiation oncologists.
The ability to combine information from multiple imaging modalities (multimodal image analysis) is a particular strength of AI. By simultaneously analyzing CT scans for anatomical detail, MRI for soft tissue contrast, and PET scans for metabolic activity, AI models can produce richer tumor characterizations than any single imaging type alone, leading to better-targeted radiation delivery.
Tumors and the organs around them change during the weeks of radiation treatment. Patients lose weight, tumors shrink in response to therapy, and anatomical landmarks shift. Traditional treatment plans are designed once at the start of treatment and fixed, meaning they may become less accurate as these changes occur. Adaptive radiation therapy (ART) addresses this by continuously updating the plan.
AI accelerates ART by rapidly analyzing daily imaging data, detecting anatomical changes, and automatically adjusting treatment parameters before each session. What would previously require hours of manual replanning can now be accomplished within the time of a single treatment appointment, making genuinely personalized day-by-day treatment practical.
Commercial systems representing this frontier include Varian's Ethos, which analyzes daily cone-beam CT images to adapt plans automatically, and Elekta's Unity system, which combines a high-field MRI scanner directly with a linear accelerator to enable real-time imaging and adaptation during treatment delivery itself.
AI can analyze historical patient data to predict how a specific patient is likely to respond to a given radiation treatment before that treatment begins. These predictions cover outcomes like tumor response rates, likelihood of local recurrence, survival probability, and the risk of specific side effects such as radiation pneumonitis or xerostomia.
Modern predictive models integrate diverse data types including imaging features (radiomics), genomic profiles, demographic information, and treatment history. By finding patterns across thousands of prior patients, machine learning models can identify which combination of features predicts the best outcome for a new patient, supporting more informed treatment selection.
Systems like IBM Watson for Oncology and RayStation already incorporate predictive analytics to support clinical decision-making. Future work aims to make these models more interpretable, allowing clinicians to understand why the model made a specific prediction rather than accepting it as a black box, which is critical for building clinical trust.
Data quality and quantity remain fundamental obstacles. AI models require large, diverse, and accurately annotated datasets to train effectively, but medical imaging data is often siloed within individual institutions, inconsistently labeled, and subject to privacy regulations that limit sharing. Models trained on limited or biased data may perform well in one hospital but poorly in another.
Interoperability is a significant practical barrier. Radiation oncology involves a complex ecosystem of imaging systems, treatment planning software, linear accelerators, and oncology information systems from different vendors, often using different data formats and communication standards. Integrating AI tools across this ecosystem requires standardization efforts that are still ongoing.
Regulatory and ethical challenges add another layer of complexity. Existing medical device regulations were not designed with continuously learning AI systems in mind, creating uncertainty about approval pathways. Ethical concerns include the risk that AI models trained on historically biased datasets could perpetuate or amplify existing healthcare disparities, and the need for AI to be explainable enough for patients and clinicians to trust its recommendations.
The review concludes that realizing AI's potential in radiation oncology requires active collaboration between researchers, clinicians, data scientists, and industry partners. Multidisciplinary teams are best positioned to build AI tools that address real clinical needs and fit naturally into existing workflows, rather than tools designed without clinical input that clinicians resist adopting.
Developing common data standards, such as DICOM for imaging data, and standardized protocols for AI model development and validation will be essential for enabling data sharing across institutions and ensuring AI tools meet consistent quality and safety benchmarks. International bodies like the IEC are already working toward such standards for AI in medical devices.
Explainable AI is highlighted as a critical priority. Clinicians are more likely to trust and use AI recommendations when they can understand the reasoning behind them. As AI models become more sophisticated, parallel investment in interpretability methods will be needed to maintain the human oversight that is both clinically and ethically necessary in high-stakes radiation therapy decisions.