Clinical Validation of AI-Assisted Radiotherapy Contouring Software for Thoracic Organs at Risk

Browse trials for

• Lung Neoplasms 7,243
• Breast Neoplasms 12,544
• Esophageal Neoplasms 1,593

Keywords

• Artificial Intelligence
• Radiotherapy
• contouring
• thoracic organs at risk

Lung Neoplasms

7,243 studies on the registry are indexed under Lung Neoplasms; 1,557 are open to participants now.

This study's enrollment of 500 is above the median of 189 across 1,514 observational studies indexed under Lung Neoplasms.

Lead sponsor

Tianjin Medical University Cancer Institute and Hospital is the lead sponsor of 484 studies on the registry; 286 are open to participants now.

Study population

Groups and cohorts

• Independent manual contouring

  Manual contouring refers to physicians using the brush tool on the contouring platform to segment thoracic organs at risk manually, without the use of auto-segmentation tools.

• AI contouring

  AI contouring refers to the auto-segmentation results generated by the Res-SE Net model, with the model integrated into the auto-segmentation software (iCurveE).

• AI-assisted contouring

  After generating the AI contouring results, investigators will import them into the contouring platform and perform manual modifications, producing the AI-assisted contouring.

Primary outcomes

1. volumetric DICE similarity coefficient, vDSC

   vDSC= 2×（A∩B）/(A+B), where A refers to the volume of the ground truth, and B refers to the volume of the manual, AI, or AI-assisted contour.

   Time frame: Within 6 months after enrollment

2. Contouring time (min)

   Manual contouring time is recorded from the time the CT is loaded on the contouring platform to the completion of contouring. AI-assisted contouring time is defined as the sum of the auto-segmentation model runtime, the transfer to the contouring platform, and the subsequent manual modification.

   Time frame: Within 6 months after enrollment

Secondary outcomes

1. 95th percentile Hausdorff Distance, HD95

   HD95(A, B) = max (h95(A, B), h95(B, A)), where h95(A, B) is the 95th percentile of the shortest distances from all points on surface A to surface B, and vice-versa for h95(B, A). A represents the ground truth and B represents the manual, AI or AI-assisted delineation

   Time frame: Within 6 months after enrollment

2. Surface DICE similarity coefficient, sDSC

   sDSC = (\|S(A) ∩ S(B)τ\| + \|S(B) ∩ S(A)τ\|) / (\|S(A)\| + \|S(B)\|), where S(A) and S(B) are the sets of points on the surfaces of A and B, S(B)τ represents the points on surface B that are within the tolerance τ of surface A, and S(A)τ represents the points on surface A that are within the tolerance τ of surface B. A represents the ground truth and B represents the manual, AI or AI-assisted delineation

   Time frame: Within 6 months after enrollment

3. Rate of time efficiency improvement

   Rate of efficiency time improvement= (manual contouring duration - AI-assisted contouring duration)/ manual contouring duration\*100%

   Time frame: Within 6 months after enrollment

4. Volumetric revision index, VRI

   VRI = \[(A- A∩B) + (B- A∩B)\] /A, where A refers to the volume of the ground truth, and B refers to the volume of the manual, AI, or AI-assisted contour.

   Time frame: Within 6 months after enrollment

5. Recall, Rec

   Rec = \| A∩B\| / A, where A refers to the volume of the ground truth, and B refers to the volume of the manual, AI, or AI-assisted contour.

   Time frame: Within 6 months after enrollment

6. Precision, Pre

   Pre= \|A∩B\| / B, where A refers to the volume of the ground truth, and B refers to the volume of manual, AI, or AI-assisted contour.

   Time frame: Within 6 months after enrollment

7. Relative volume difference, RVD

   RVD = \|A-B\| /A, where A refers to the volume of the ground truth, and B refers to the volume of the manual, AI, or AI-assisted contour.

   Time frame: Within 6 months after enrollment

8. Investigators satisfaction score for AI contouring

   Evaluated on a 1-5 Likert scale: 1 - strongly dissatisfied, 2 - dissatisfied, 3 - neutral, 4 - satisfied, 5 - strongly satisfied.

   Time frame: Within 6 months after enrollment

Other outcomes

1. Number of adverse events, AEs

   Participant Adverse events during CT scanning

   Time frame: Within 1 day after CT scanning

2. Number of device defects during AI-assisted contouring

   Number of failures in generating, transferring, or saving auto-segmentation results

   Time frame: Within 6 months after enrollment

Publications

• Niu G, Guan Y, Zhang Y, Song Y, Yan M, Li S, Liu T, Huang S, Chen J, Wang X, Zhang W, Meng M, Liu Y, Chen J, Fu Y, Zhao D, Huang J, Yang K, Cao J, Yuan H, Guo S, Pei X, Wu D, Nan Y, Yan Z, Lu Y, Zhao L, Yuan Z. A prospective multicenter trial of deep learning auto-segmentation for organs at risk in thoracic radiotherapy. Nat Commun. 2026 Mar 31;17(1):4633. doi: 10.1038/s41467-026-70863-9. PubMed 41917034 ↗

Individual participant data

Plan to share: Yes — The protocol of this study are available from the corresponding author upon reasonable request after the manuscript publication.

Supporting information: Study protocol