KEY POINTS
- This collaborative study involving Leiden University Medical Center, Holland Proton Therapy Centre, and partner institutions compared conventional manual contour editing with an AI-assisted pencil tool for correcting automatically generated head and neck tumor contours on registered PET/CT images. The user study involved four radiation oncologists with 2, 4, 11, and 21 years of contouring experience, plus seven nonexpert AI researchers, each working on six patient cases using both editing methods.
- The underlying auto-contouring and refinement models were trained using 452 paired PET/CT scans from the HECKTOR2022 dataset, originating from centers in Canada, Switzerland, and the United States, with an independent geographic test set of 72 scans from France. Both models used a relatively lightweight 3D U-Net architecture containing approximately 1.2 million parameters. For contour refinement, the AI tool combined CT, PET, the existing contour, and user-drawn positive or negative scribbles, using sparse two-dimensional interactions to update the full three-dimensional tumor volume.
- The AI pencil substantially reduced editing time for experienced radiation oncologists. Across 24 expert editing sessions, the AI-assisted approach was 16–97% faster than the manual brush, with a median time saving of 70%, equivalent to 234 seconds per session. Experts also required 55–97% fewer editing interactions, depending on the patient and observer. The model itself had a mean inference time of approximately 2.9 seconds, included in the reported total editing time.
- Time savings were also observed among nonexpert users, although the magnitude was smaller and more variable. Across 42 nonexpert sessions, AI-assisted correction was 5–78% faster than manual editing, with a median saving of 49%, or 258.5 seconds. The number of editing interactions decreased by 14–87%, while the number of manually drawn pixels decreased by 63–93% in nonexpert sessions and 81–99% in expert sessions. Unlike the clinical experts, nonexperts were shown the reference contours and specifically asked to reproduce them.
- Faster correction did not produce an obvious reduction in final contour accuracy. The initial auto-contouring model achieved a mean Dice score of 0.76 and surface Dice at 2 mm of 0.72. After refinement by clinical experts, final contours achieved mean Dice 0.84 with both the manual brush and AI pencil when compared with the reviewed dataset reference. Corresponding surface Dice values were 0.81 with manual editing versus 0.85 with the AI pencil. These results support comparable geometric accuracy, although no formal clinical equivalence margin was predefined.
- Interobserver consistency was also similar between editing approaches. Mean expert interobserver Dice and surface Dice were 0.89 and 0.90 with the manual brush, compared with 0.90 and 0.92 using the AI pencil. Agreement between the final contours produced by the two tools averaged Dice 0.81 and surface Dice 0.77. Qualitative analysis showed that a single scribble could correct errors across neighboring slices, but some interactions produced incomplete or temporarily worse segmentations, particularly when the initial auto-contour was already relatively accurate.
- The study's applicability is limited by its six selected patient cases, relatively small expert cohort, and use of an experimental web-based interface rather than a fully commissioned clinical treatment-planning workflow. Experts did not have access to the full clinical information ordinarily used for target delineation, including MRI, endoscopic findings, and clinical notes. The investigation also did not evaluate downstream treatment-plan dosimetry, clinical safety, integration with commercial systems, or performance across a broader range of anatomical and segmentation failures.
CLINICAL TAKEAWAY
AI-assisted contour refinement using sparse user interactions may substantially reduce the time required for radiation oncologists to correct head and neck tumor auto-contours without materially compromising final geometric accuracy. However, these findings are based on a limited experimental user study, and broader validation within routine contouring and treatment-planning workflows is needed before the approach can replace conventional manual correction tools.