An evaluation of rectum contours generated by artificial intelligence automatic contouring software using geometry,...

Area of Science:

• Radiation Oncology
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Accurate delineation of organs at risk, such as the rectum, is crucial for safe and effective radiotherapy planning.
• Deep learning (DL) models offer potential for automating contour generation, but their clinical utility requires rigorous validation against expert delineations.

Purpose of the Study:

• To evaluate the geometric and dosimetric accuracy of rectum organ at risk contours generated by a commercial DL auto-contouring model compared to clinician-drawn contours.
• To assess the impact of contour differences on toxicity modeling for predicting late rectal bleeding in prostate cancer patients undergoing 3D-conformal radiotherapy.

Main Methods:

• Retrospective analysis of 308 prostate cancer patients treated with 3D-conformal radiotherapy.
• Rectum contours were generated using a commercial DL auto-contouring software (Limbus Contour v1.8.0b3) without manual editing.
• Comparison of auto-contours and clinician contours using Dice Similarity Coefficient (DSC), Hausdorff Distance (HD), volume difference, and dose-volume histogram (DVH) metrics. Toxicity prediction models incorporated Equivalent Uniform Dose (EUD) and clinical factors.

Main Results:

• Median DSC was 0.85, median HD was 1.38 mm, and median volume difference was -1.73 cc, indicating good overlap and proximity.
• Median DVH differences were small (<1.5%), but interquartile ranges up to 8.0% highlight inter-patient variability.
• Toxicity models incorporating DL-generated contours showed comparable predictive accuracy (AUC 0.66) for rectal bleeding as those using clinician contours (AUC 0.65).

Conclusions:

• While DL auto-contouring demonstrates geometric and dosimetric similarity to clinician contours, the observed inter-patient variability necessitates continued clinician oversight in radiotherapy planning.
• The DL auto-contouring model shows potential for integration into supervised workflows, offering efficiency without compromising predictive accuracy for rectal bleeding in this cohort.