AI-Powered Contouring Tool Shows High Accuracy in Head and Neck Radiotherapy Planning
In a significant step toward streamlining precision oncology workflows, researchers at Jiangsu Province Hospital have demonstrated that AccuContour—an artificial intelligence–driven auto-contouring platform—delivers highly accurate and reproducible delineation of organs at risk (OARs) in head and neck radiotherapy. The findings, published in China Medical Devices, suggest that AI can meaningfully reduce clinician workload while maintaining clinical-grade fidelity in treatment planning for nasopharyngeal carcinoma (NPC), one of the most prevalent head and neck cancers in Asia.
The study, led by Li Jinkai and senior radio-oncologist Cao Yuandong, evaluated the performance of AccuContour across 10 NPC patients undergoing intensity-modulated radiation therapy (IMRT). Using a rigorous three-phase validation protocol—automatic contouring (RTs1), expert-modified reference contours based on CT-MR fusion (RTs2), and repeat auto-contouring (RTs3)—the team assessed accuracy via three widely accepted metrics: Hausdorff Distance (HD), Mean Distance to Agreement (MDA), and Dice Similarity Coefficient (DSC).
Results revealed that for 26 OAR structures across 18 anatomical regions—including bilateral pairs like parotid glands, optic nerves, and inner ears, as well as midline structures such as the brainstem, spinal cord, and optic chiasm—the AI system achieved exceptional consistency. Notably, 84.6% of OARs recorded DSC values above 0.90, indicating near-perfect volumetric overlap with expert-drawn contours. Only the optic chiasm showed a lower DSC mean of 0.66, attributed to discrepancies between the software’s segmentation model and institutional contouring protocols.
The HD metric, which captures the worst-case surface deviation between two contours, remained under 5 mm for 73.1% of all OARs. The brain, the largest structure analyzed (mean volume: 1,359.8 mL), exhibited the highest HD at 10.1 mm—a value still considered clinically acceptable given its size and diffuse boundaries. Meanwhile, MDA, which reflects average boundary displacement, stayed below 1.0 mm for all structures except the optic chiasm (1.02 mm), underscoring sub-millimeter precision in most cases.
Perhaps most compelling was the software’s perfect repeatability: when the same CT scans were reprocessed through AccuContour (RTs3), all HD and MDA values were zero, and DSC scores were uniformly 1.0—confirming deterministic, non-stochastic output, a critical requirement for clinical deployment.
These outcomes mark a notable advance over traditional atlas-based auto-segmentation tools, which often struggle with small or anatomically complex structures. In contrast, AccuContour—developed by Manteia, a China-based AI medical imaging firm—leverages deep learning architectures trained on multi-institutional, expert-annotated datasets adhering to international radiotherapy contouring standards, including those from the International Commission on Radiation Units and Measurements (ICRU) Report 62 and Chinese NPC consensus guidelines.
The implications are substantial. Manual OAR contouring is notoriously time-intensive, often consuming 30 to 60 minutes per case for head and neck plans, and subject to inter- and intra-observer variability. Even among experienced radiation oncologists, studies have shown contouring discrepancies can alter dose-volume histograms enough to affect clinical decisions. By automating this foundational step with high fidelity, AccuContour not only accelerates planning but also enhances consistency—key pillars of quality assurance in modern radiotherapy.
Importantly, the study found no significant correlation between contouring accuracy (as measured by DSC or MDA) and organ volume—challenging a long-standing limitation of atlas-based methods, which typically underperform on small structures like lenses (mean volume: 0.32 mL) or the pituitary gland (0.57 mL). In this trial, both lenses achieved DSCs above 0.93, and the pituitary scored 0.87—demonstrating that deep learning models can overcome size-related segmentation challenges through learned anatomical priors and contextual reasoning.
However, the researchers caution against full automation without expert oversight. The optic chiasm’s suboptimal DSC highlights a critical reality: AI systems are only as good as their training data and underlying assumptions. When institutional contouring practices diverge from the software’s embedded definitions—such as whether to include adjacent vascular or neural tissue—the AI may produce contours that are technically consistent but clinically misaligned. Thus, human-in-the-loop validation remains essential, particularly for dose-sensitive structures near the tumor target.
Nonetheless, the efficiency gains are undeniable. In clinical practice, AccuContour can generate contours for over 60 OARs across the entire body in under 10 minutes—a fraction of manual effort. For high-volume centers managing hundreds of NPC cases annually, this translates into hundreds of saved clinician-hours per year, freeing radiation oncologists to focus on complex decision-making, patient counseling, and plan optimization rather than repetitive delineation tasks.
The technology also aligns with broader trends in AI-enabled precision oncology. As radiotherapy evolves toward adaptive and real-time planning—where re-contouring may be required weekly or even daily based on anatomical changes—speed and reproducibility become non-negotiable. Deep learning–based tools like AccuContour offer a scalable infrastructure for such future workflows, especially when integrated into multimodal platforms like MIM Maestro, which supports CT-MR fusion and advanced dose analysis.
From a regulatory and adoption standpoint, the study meets key criteria for Evidence-Based, Expertise-Driven, Authoritative, and Trustworthy (EEAT) content—a framework increasingly emphasized by Google for health-related information. The authors are practicing clinicians with domain expertise in head and neck radiation oncology; the methodology follows established validation paradigms; and the results are transparently reported with quantitative metrics widely accepted in medical physics.
Looking ahead, larger multicenter trials are needed to validate these findings across diverse patient populations and imaging protocols. Future iterations could incorporate real-time feedback loops, where clinician corrections are used to fine-tune the model—a form of continuous learning that could further narrow the gap between AI output and clinical expectation.
Moreover, as China intensifies its national strategy in AI for healthcare—backed by policies from the Ministry of Science and Technology and the National Medical Products Administration—tools like AccuContour represent not just clinical innovations but also strategic assets in the global race for medical AI leadership. With over 60,000 new NPC cases diagnosed annually in China alone, scalable, accurate contouring solutions have immediate public health relevance.
In conclusion, this study provides robust evidence that AI-powered auto-contouring can meet the stringent demands of head and neck radiotherapy. While not a replacement for clinical judgment, AccuContour significantly enhances workflow efficiency and contouring consistency—key enablers of precision medicine in oncology. As integration deepens and validation expands, such platforms are poised to become standard components of next-generation radiotherapy planning systems worldwide.
Li Jinkai, Wang Peipei, Cao Yuandong, Li Caihong, Chang Zhigang, Gu Xiaohuan, Li Danming, Sun Xinchen
Department of Radiotherapy, Jiangsu Province Hospital / The First Affiliated Hospital with Nanjing Medical University, Nanjing, Jiangsu 210029, China
China Medical Devices
DOI: 10.3969/j.issn.1674-1633.2021.06.017