After surgery for endometrial cancer, many patients receive brachytherapy - a form of radiation therapy where a small radioactive source is placed directly inside the body (in this case, inside the vagina) to irradiate any remaining cancer cells in the vaginal area. Unlike external radiation beams, brachytherapy delivers a very concentrated dose very close to the treatment site, which helps protect surrounding healthy organs.
Before each brachytherapy treatment fraction, a radiation oncologist must carefully draw outlines - called contours - on a CT scan to delineate both the treatment target (known as the high-risk clinical target volume, or HR CTV) and the organs at risk (OARs, specifically the rectum and urethra) that need to be protected from excessive radiation. This contouring process is critically important: too little coverage means residual cancer cells may survive, while too much coverage means healthy organ damage. The process currently takes 10 to 26 minutes per session by hand.
A 2024 study published in the Journal of Applied Clinical Medical Physics developed and tested a new deep learning model - called 3D Prompt-nnUnet - to automate this contouring task. The goal was to match the accuracy of experienced oncologists while reducing the time from more than 20 minutes to just a few seconds per structure.
The model builds on an established deep learning architecture called nnUnet, which automatically configures itself for medical image segmentation tasks without requiring manual tuning of settings. The researchers enhanced this with a novel prompt-based component inspired by the Segment Anything Model (SAM) - a powerful general-purpose image segmentation system developed by Meta AI. The key insight is that providing the model with a simple bounding box around the structure of interest dramatically improves accuracy compared to blind segmentation.
The model comes in two modes: Label-Prompt (LP) uses a bounding box derived from actual manual labels (representing the case where a physician marks the approximate location of the target on a few 2D slices), and Predict-Prompt (PP) uses a bounding box automatically predicted by the nnUnet model itself. Technically, this bounding box is encoded as coordinate information via a cross-attention mechanism - a mathematical approach borrowed from large language models - that allows the model to focus its attention on the correct anatomical region.
Training used 321 CT scans for HR CTV segmentation and 125 CT scans for OAR segmentation. Each CT scan corresponded to a single brachytherapy treatment fraction, with patients receiving 3-6 fractions each. The ground truth contours used for training and evaluation were drawn by senior radiation oncologists and verified by a second senior oncologist. Three-fold cross-validation was used to test stability across different data splits.
The LP version of Prompt-nnUnet delivered the best results across all structures. For the HR CTV (the main treatment target), it achieved a mean Dice Similarity Coefficient (DSC) of 0.96 - meaning 96% overlap with the expert-drawn contour. The Hausdorff Distance (HD95%) was 1.66 mm, meaning the furthest boundary deviation fell within about 1.7 mm. For the rectum, DSC was 0.91 with HD95% of 3.07 mm, and for the urethra, DSC was 0.83 with HD95% of 1.35 mm.
In head-to-head comparison with human oncologists of varying experience, the LP model performed comparably to senior oncologists (10 or more years of experience) for HR CTV (DSC 0.96 vs. 0.90 for senior oncologists), and outperformed both junior (0-5 years, DSC 0.82) and intermediate (5-10 years, DSC 0.87) oncologists. For the rectum and urethra, the model matched senior oncologist performance. The model was immediately usable without modification in 53.3% of HR CTV cases, with only minor adjustments needed in the remaining 46.7% - and no cases requiring major revision.
In terms of speed, the Prompt-nnUnet completed HR CTV segmentation in 1.71 seconds and OAR segmentation in 1.51 seconds. Manual contouring by junior oncologists took 22.8 minutes, intermediate oncologists 17.6 minutes, and senior oncologists 11.6 minutes. Even the baseline nnUnet model required 33 seconds per structure. The LP model thus achieves an approximately 400-fold speed improvement over manual contouring by experienced oncologists.
Accurate contouring is only valuable if it translates into accurate radiation dose delivery. The researchers therefore evaluated the dosimetric impact - comparing the radiation dose parameters calculated from AI-generated contours versus expert-drawn contours. This analysis is critical because even geometrically similar contours can lead to meaningfully different dose distributions depending on where boundaries fall relative to the radiation source.
For HR CTV, the LP model showed smaller differences in dose coverage parameters compared to both PP and standard nnUnet. Specifically, the deviation in D90% (the dose delivered to 90% of the target volume) was 23 cGy for LP versus 44-45 cGy for PP and standard nnUnet. This means that when treatment plans are calculated from LP contours, the resulting dose distributions more closely match what was intended based on the expert-drawn plan.
For OARs, LP similarly showed smaller deviations in dose parameters for both the rectum and urethra. This has direct clinical implications: using the LP model rather than the standard nnUnet or manual contouring by junior or intermediate oncologists could lead to treatment plans that more accurately deliver the prescribed dose to the cancer tissue while better protecting healthy organs from radiation damage.
This study demonstrates that AI-assisted contouring can reach senior-oncologist-level accuracy in brachytherapy treatment planning for endometrial cancer. The time savings are dramatic - from 10-26 minutes down to under 2 seconds per structure - and the model could be particularly valuable in settings where experienced radiation oncologists are limited in number. By enabling junior and intermediate oncologists to produce contours of senior-level quality with a few simple steps (marking the target location on 2-3 CT slices to generate the bounding box), the model also democratizes access to high-quality treatment planning.
The study also highlights the promise of prompt-guided deep learning in medical imaging more broadly. The key insight - that providing a simple location hint to a neural network dramatically improves its ability to delineate complex anatomical structures - has implications beyond brachytherapy. Similar prompt-based approaches may improve AI segmentation in radiotherapy planning for other cancer types and anatomical sites.
Important limitations include single-center data (limiting generalizability), challenges in complex anatomy like the urethra, and the relatively small OAR dataset (125 cases). One illustrative case from related work showed the model briefly misidentifying a prosthetic implant as tumor tissue - highlighting the need for physician oversight of AI-generated contours before clinical use. Validation in larger multi-center studies is the critical next step before this approach is deployed routinely in clinical practice.