Evaluation of Model Performance and Clinical Usefulness in Automated Rectal Segmentation in CT for Prostate and...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiotherapy

Background:

• Precise rectal delineation is critical for pelvic cancer treatment planning.
• Manual segmentation is time-consuming and prone to inter-observer variability.
• Sex-specific anatomical differences can enhance segmentation accuracy.

Purpose of the Study:

• To develop and validate a sex-aware deep learning pipeline for automated rectal segmentation in CT scans.
• To leverage patient sex classification to improve segmentation performance for prostate and cervical cancers.
• To assess the clinical utility and efficiency of the AI-driven segmentation approach.

Main Methods:

• A two-stage deep learning pipeline using CT scans from 186 patients.
• Stage 1: CNN model for automated patient sex classification.
• Stage 2: Sex-aware U-Net model for automated rectal segmentation, incorporating sex-specific features.

Main Results:

• Sex-classification model achieved 94.6% accuracy (AUC = 0.98).
• Improved anatomical consistency in segmentation outputs.
• High Dice Similarity Coefficient (DSC) values (0.91 for prostate, 0.89 for cervical).
• Clinically acceptable surface distance metrics (HD: 3.4 ± 0.8 mm, ASD: 1.2 ± 0.3 mm).
• Reduced contouring time from 12.7 min to 4.3 min.
• 89.2% of automated contours rated as excellent by a radiation oncologist.

Conclusions:

• The sex-aware deep learning framework provides accurate and robust rectal segmentation in pelvic CT imaging.
• Explicitly modeling sex-specific anatomical differences enhances segmentation performance.
• This AI approach improves contouring efficiency and clinical consistency in radiotherapy workflows.