https://atlas.rsna.org/schemas/2025-11/model.json

AI-assisted Contour Editing (AIACE)

https://pubmed.ncbi.nlm.nih.gov/36204538/

1.0

Steve Jiang, Department of Radiation Oncology, University of Texas Southwestern Medical Center; email: ude.nretsewhtuoSTU@gnaiJ.evetS

For the UTSW test dataset: IRB-approved and de-identified; informed consent waived because data were collected retrospectively with minimal risk.

U-Net–based deep convolutional neural network for interactive contour editing (three-channel input: original CT slice, current segmentation mask, and click image; output: updated mask).

Not stated.

Assists clinicians in efficiently and effectively editing organs-at-risk contours for radiotherapy planning, reducing manual editing time and improving contour quality, including in online adaptive radiotherapy workflows.

Clinical decision support and workflow optimization for contour editing during treatment planning/online adaptive radiotherapy.

For failure analysis, an example quality threshold was HD95 < 2.5 mm within 20 clicks.

Interactive assistance; clinician-in-the-loop with iterative clicks guiding automated contour updates.

Editing of head-and-neck organs-at-risk contours on axial CT images for radiotherapy planning; intended for use by clinicians in radiation oncology settings.

2D axial head-and-neck CT image slice; current segmentation mask; clinician-provided click image (Gaussian around clicked boundary point).

Start from an automatically generated initial contour; at each iteration, the clinician clicks on the boundary location with the largest perceived error. The click is converted to a Gaussian point in a click image (radius ~10 pixels). Provide the three-channel input (image, current mask, click image) to the model to obtain an updated contour. Repeat until clinically acceptable.

Proof-of-concept using 2D axial images only; clinician interactions simulated during training/testing; performance reported with simulated clicks placed at largest-error boundary points (upper bound of performance); occasional failures when successive clicks provide insufficient new information; model trained and evaluated on head-and-neck CT data; generalization beyond studied organs/modalities not demonstrated.

Use downstream of an automatic segmentation model to iteratively correct contours with minimal clicks; not a replacement for fully automatic segmentation.

Five independently trained models with different random seeds showed consistent improvements across datasets; datasets and parameter settings described; training/validation data sources publicly available.

Real-time interaction feasible; ~20 ms per iteration on a single NVIDIA GeForce Titan X GPU.