Residency · Residency · Radiation Oncology
Artificial Intelligence in Radiation Oncology: Auto-Contouring and Treatment Planning
Introduction
Artificial intelligence (AI), especially deep learning, is revolutionizing radiation oncology by automating the labor-intensive steps involved in treatment planning. The two areas where AI has the most immediate clinical impact are auto-contouring, which involves the automated delineation of target volumes and organs at risk (OARs), and automated treatment planning, which includes knowledge-based planning and deep learning-based plan generation. This lecture explores the current state of AI applications in radiation oncology, their clinical validation, and future directions.
Foundations of AI in Radiation Oncology
Machine Learning vs. Deep Learning
Machine learning refers to algorithms that learn patterns from data without explicit programming. This category includes methods such as random forests, support vector machines, and gradient boosting. Deep learning is a subset of machine learning that employs multi-layer neural networks, including convolutional neural networks (CNNs) and transformers. Deep learning has demonstrated superior performance for image-based tasks such as contouring, image registration, and outcome prediction. However, it requires large, annotated training datasets to achieve high accuracy.
Convolutional Neural Networks (CNNs)
CNNs are neural network architectures specifically designed for image processing, capable of learning hierarchical spatial features from input images. The U-Net architecture, which features an encoder-decoder structure with skip connections, has become the standard for medical image segmentation. Variants such as 3D U-Net, nnU-Net, and attention U-Net are commonly used for volumetric medical image segmentation. Training these networks requires paired imaging data with expert-delineated contours serving as ground truth.
Transformer-Based Models
Originally developed for natural language processing, transformer-based models have been adapted for image segmentation tasks. Vision Transformers (ViTs) and hybrid CNN-Transformer architectures can capture long-range spatial dependencies more effectively than pure CNNs. These models are emerging as state-of-the-art solutions for complex segmentation challenges in medical imaging.
Auto-Contouring
Clinical Need
Manual contouring is a time-consuming process, often requiring 30 to 120 minutes per patient for complex cases. Additionally, there is significant inter-observer variability, with contour differences ranging from 20% to 50% for some structures. Auto-contouring can reduce this time to 5 to 15 minutes, allowing for review and editing rather than full manual delineation. This automation improves consistency and enhances workflow efficiency.
OAR Auto-Contouring
Auto-contouring of organs at risk is the most mature clinical application of AI, with several FDA-cleared products currently available. OAR contouring is well-suited to AI because of the consistent anatomy of these structures. Performance metrics commonly used include the Dice similarity coefficient (DSC), which measures the overlap between the auto-contour and expert contour, with a value of 1.0 indicating perfect agreement. Typical DSC values for OARs are 0.85 to 0.90 for parotid glands, greater than 0.95 for lungs, above 0.90 for the heart, and between 0.80 and 0.90 for the brainstem. Other metrics include the Hausdorff distance, which measures the maximum distance between contour surfaces, and the mean surface distance, which averages the distances between surfaces.
Commercial AI Auto-Contouring Systems
Several FDA-cleared commercial systems exist, including Limbus AI, Radformation, MVision AI, Therapanacea, and Siemens AI-Rad. The typical workflow involves importing a CT scan, generating auto-contours within 1 to 5 minutes, and then having a physician review and edit the contours as needed. These systems have been validated across multiple anatomical sites such as the head and neck, thorax, abdomen, pelvis, and brain. Ongoing improvements are driven by expanding training datasets and model updates.
Target Volume Auto-Contouring
Auto-contouring of target volumes presents greater challenges than OAR contouring because it requires clinical judgment. While AI can assist with gross tumor volume (GTV) delineation for visible tumors, physician oversight remains essential. Clinical target volume (CTV) delineation is more complex, as it involves understanding patterns of microscopic disease spread. This area is an active research focus, with promising initial results. AI models are beginning to incorporate lymph node CTV delineation guidelines, such as head and neck nodal levels, to improve accuracy.
Automated Treatment Planning
Knowledge-Based Planning (KBP)
Knowledge-based planning systems, such as RapidPlan by Varian, utilize libraries of prior high-quality treatment plans to predict achievable dose-volume histograms (DVHs) for new patients. These models are trained on institutional plan databases and use geometric features of patient anatomy to predict DVH parameters. The system then automatically generates optimization objectives, reducing planning time and inter-planner variability. KBP has been validated across multiple disease sites, including prostate, head and neck, lung, and breast cancers.
Deep Learning-Based Dose Prediction
Deep learning models have been developed to predict three-dimensional dose distributions directly from CT images and contour data. The input consists of patient anatomy, including CT scans and contours, and the output is a predicted dose distribution. These models can generate clinical-quality plans within minutes. Often, a cascade workflow is employed, where the predicted dose distribution serves as the optimization objective for generating a deliverable treatment plan. The accuracy of these models approaches that of expert human planners in many disease sites.
Fully Automated Planning
End-to-end automated planning systems take CT images and contours as input and produce deliverable treatment plans without manual intervention. Auto-planning modules are integrated into commercial treatment planning systems (TPS), such as Pinnacle Auto-Planning and Eclipse scripts. These systems reduce physicist planning time from hours to minutes. Despite automation, quality assurance of AI-generated plans remains essential to ensure safety and efficacy.
Clinical Validation
Multiple studies have demonstrated that AI-generated treatment plans are clinically acceptable in 70% to 95% of cases without modification. The dosimetric quality of these plans is comparable to or better than manual planning. Prospective clinical implementation studies are ongoing. A key consideration is that all plans must meet institutional standards and undergo physicist review before clinical use.
Quality Assurance and Validation
Validation Metrics
Validation of AI-generated contours and plans employs several metrics. Geometric metrics include the Dice coefficient, Hausdorff distance, and surface distance for evaluating auto-contours. Dosimetric metrics assess target coverage (such as D95 and V100%), adherence to OAR constraints, conformity index, and gradient index. Clinical acceptability is determined by physician assessment of contour adequacy and plan quality. Blinded comparison studies between AI and manual methods have demonstrated non-inferiority in many clinical settings.
Failure Modes and Limitations
AI systems may fail in out-of-distribution cases, such as patients with unusual anatomy, imaging artifacts, or rare tumor presentations. Metal artifacts from hip prostheses or dental implants can degrade auto-contouring accuracy. Prior surgery or significant anatomic alterations may not be well-represented in training datasets, leading to errors. Therefore, physician review remains mandatory; AI serves as a tool to support, not replace, clinical judgment.
Regulatory Considerations
In the United States, commercial AI contouring tools require FDA 510(k) clearance, while in the European Union, CE marking is necessary. Post-market surveillance and failure reporting are important components of regulatory compliance. Institutions must perform commissioning and validation of AI tools before clinical deployment to ensure safety and effectiveness.
Other AI Applications in Radiation Oncology
Outcome Prediction
Radiomics involves extracting quantitative imaging features from CT, MRI, or PET scans. Machine learning models use these features to predict treatment response, toxicity, and survival outcomes. Delta radiomics, which analyzes changes in imaging features during treatment, adds dynamic information. Integrating radiomics with genomic and clinical data enables comprehensive prognostic modeling.
Adaptive Radiation Therapy
AI facilitates daily plan adaptation by accounting for anatomical changes during treatment. Automated re-contouring on daily cone-beam CT (CBCT) or MRI allows real-time dose accumulation and plan modification. This capability is essential for MR-LINAC workflows, such as those using Unity or ViewRay systems.
Image Synthesis
Deep learning models can generate synthetic CT images from MRI scans, supporting MRI-only planning workflows. These pseudo-CT images enable accurate dose calculation without the need for a separate CT simulation. This approach has been validated for brain, head and neck, and pelvic regions.
Natural Language Processing
Natural language processing techniques automate the extraction of clinical information from radiation oncology notes. This supports quality metric tracking, clinical trial eligibility screening, and automated generation of treatment summaries.
Future Directions
Foundation Models
Large pre-trained models, known as foundation models, are being adapted for radiation oncology tasks. The Segment Anything Model (SAM) and its medical adaptations aim to provide universal segmentation capabilities. These models have the potential to reduce training data requirements through transfer learning.
Personalized Treatment
AI-driven dose painting based on functional imaging modalities such as PET and fMRI is an emerging area. Treatment optimization incorporating biomarkers promises more personalized therapy. Real-time adaptive therapy supported by AI-driven decision tools represents a future advancement.
Clinical Integration Challenges
Integrating AI into clinical workflows without disrupting efficiency remains a challenge. Building trust and interpretability of AI decisions is critical. Data privacy concerns motivate the use of federated learning approaches. Detecting and mitigating bias in training datasets is essential to ensure equitable AI performance.
Key Clinical Pearls
AI auto-contouring is currently the most clinically mature AI application in radiation oncology, with multiple FDA-cleared products achieving expert-level performance for OAR delineation. This technology reduces planning time from hours to minutes and improves inter-observer consistency; however, physician review and editing remain essential. Knowledge-based planning systems like RapidPlan and deep learning dose prediction models can generate clinically acceptable plans in 70% to 95% of cases. AI should be regarded as a decision-support tool rather than a replacement for clinical expertise. Failure modes include unusual anatomy, imaging artifacts, and out-of-distribution cases. The integration of AI into adaptive radiation therapy workflows, such as MR-LINAC and daily adaptation, represents the next frontier in clinical implementation.
References
- Cardenas CE, Yang J, Anderson BM, et al. Advances in auto-segmentation. Semin Radiat Oncol. 2019;29(3):185-197.
- Netherton TJ, Cardenas CE, Rhee DJ, et al. The emergence of artificial intelligence within radiation oncology treatment planning. Oncology. 2021;35(3):124-134.
- Vandewinckele L, Claessens M, Dinkla A, et al. Overview of artificial intelligence-based applications in radiotherapy: recommendations for implementation and quality assurance. Radiother Oncol. 2020;153:55-66.
- Shiraishi S, Moore KL. Knowledge-based prediction of three-dimensional dose distributions for external beam radiotherapy. Med Phys. 2016;43(1):378-387.