Deep learning automates organ contouring for brachytherapy—works even with interstitial needles in place.

Deep learning automates organ contouring for brachytherapy—works even with interstitial needles in place.

Three-dimensional image-guided brachytherapy (3D-IGBT) for cervical and endometrial cancer requires precise contouring of organs at risk (OARs: bladder, small bowel, rectum, sigmoid) for treatment planning. Manual contouring is time-consuming and variable. This study developed and tested a deep learning model (nnU-Net) on 140 brachytherapy cases, including 74 with interstitial needles.

Mean Dice similarity coefficients ranged from 0.76 (sigmoid) to 0.96 (bladder). Mean processing time was 30.3 seconds. Crucially, accuracy was not significantly different in cases with vs. without interstitial needles—a key practical advantage.

The model shows promise for clinical workflow improvement in gynecologic brachytherapy.

Key Findings

• nnU-Net achieved Dice of 0.96 (bladder), 0.83 (rectum), 0.79 (small bowel), 0.76 (sigmoid)
• Mean processing time: 30.3 seconds per case
• No significant accuracy difference between cases with and without interstitial needles
• Dosimetric impact (ΔD2cc) was clinically acceptable for all OARs
• Model developed on 100 patients, tested on 60 cases across two institutions

Implications

Automated OAR contouring in gynecologic brachytherapy could dramatically reduce planning time and inter-observer variability, potentially increasing access to high-quality brachytherapy, particularly in high-volume centers. The needle-invariant performance is clinically important.

Caveats

Single-institution development data; abstract-only. Performance metrics reflect geometric accuracy but not full clinical workflow validation. External multicenter validation needed. Sigmoid colon Dice (0.76) may be insufficient for some clinical applications.

Source: Journal of applied clinical medical physics — 2026-04-01