Prostate cancer is the second most common cancer in men worldwide, and radiation therapy (RT) is a cornerstone of its treatment. The goal of RT is to deliver a high, precisely shaped radiation dose to the tumor while protecting nearby healthy tissue from damage.
Before treatment can begin, a radiation oncologist must carefully draw boundaries around the target volumes (prostate, seminal vesicles, lymph nodes) and organs at risk (OARs) such as the bladder, rectum, and femoral heads on CT scan images. This manual contouring process is time-consuming and introduces inter-observer variability -- different clinicians often draw slightly different boundaries.
Modern delivery techniques like Intensity-Modulated Radiation Therapy (IMRT) and Volumetric-Modulated Arc Therapy (VMAT) create very sharp dose gradients, meaning even small contouring errors can lead to the tumor receiving too little dose or healthy organs receiving too much. Accurate delineation is therefore critically important.
Deep learning-based auto-segmentation has emerged as a promising solution. By learning from many expert-contoured cases, these algorithms can automatically generate contours that closely match what a specialist would draw, in a fraction of the time.
This retrospective study included 118 prostate cancer patients treated with IMRT at a hospital in Tehran, Iran. CT images were acquired with a 16-slice scanner at 110 kV with 3-mm slice thickness. The dataset was split 80/20 into training (95 patients) and testing (23 patients) sets.
The researchers applied the nnU-Net (no new U-Net) framework -- a self-configuring deep learning architecture that automatically adapts to the characteristics of any medical imaging dataset. It generates three architecture variants (2D, 3D full-resolution, and 3D cascade) and selects the best configuration, eliminating the need for extensive manual tuning.
The model was trained to simultaneously segment seven structures: prostate, seminal vesicles, lymph nodes, rectum, bladder, and both femoral heads. Training used a Dice coefficient loss function with the Adam optimizer, running for 150 epochs with a batch size of 4. Total training time was approximately 11 hours on a standard desktop GPU.
After segmentation, several post-processing steps refined the output: morphological closing operations filled gaps in segmented regions, connected component analysis isolated each organ from background, and smoothing reduced sharp edge artifacts. These steps were especially important for complex structures like the prostate and seminal vesicles.
The Dice Similarity Coefficient (DSC) -- measuring overlap between automatic and manual contours on a 0-to-1 scale -- showed the best results for the bladder (DSC 0.97) and femoral heads (DSC 0.96 for both sides). These structures have simple geometry and high contrast on CT images, making them easiest to segment accurately.
The rectum achieved a DSC of 0.90, while the prostate reached 0.82 and lymph nodes 0.77. The seminal vesicles had the lowest performance at DSC 0.69, reflecting their small size, complex shape, and high anatomical variation between patients. Segmentation of the entire CT volume took approximately 100 seconds per patient, compared to 20-30 minutes for manual contouring.
When compared to other published methods in a benchmark table, the nnU-Net achieved performance comparable to or better than established models, including the 3D U-Net (bladder DSC 0.97, rectum 0.89, prostate 0.87) and GAN-based approaches (prostate DSC 0.91). The nnU-Net was notable for being the only model to simultaneously segment all seven structures including seminal vesicles and lymph nodes.
Beyond geometric accuracy, the researchers took an important additional step: generating full IMRT treatment plans based on the automatic contours and comparing the resulting dose distributions to plans based on manual contours. This is clinically essential because a geometrically imperfect contour might still produce an acceptable treatment plan.
For the primary target (prostate) and OARs (rectum, bladder), no statistically significant differences in dose-volume histogram (DVH) parameters were found between plans using automatic versus manual contours. This suggests the segmentation quality is clinically acceptable for these structures.
Statistically significant differences were observed for the pelvic lymph nodes and femoral heads. For lymph nodes, this likely reflects the greater difficulty in delineating this diffuse, low-contrast structure. For femoral heads, even minor contouring differences affected the measured dose to small volumes within these structures, though the clinical significance of this difference is likely minimal.
A key strength of this study is that it evaluated the model using both geometric metrics (DSC, Hausdorff distance) and clinically meaningful dosimetric endpoints. Most published studies rely only on geometric metrics, which do not directly translate to treatment plan quality. The high agreement in DVH parameters supports the clinical readiness of this approach.
The nnU-Net's self-adapting nature means it requires no manual hyperparameter tuning for new datasets, streamlining the development workflow. Its cascade architecture -- where a first pass generates coarse segmentations that are refined by a second pass -- allows it to handle structures of very different sizes in a single model.
Limitations include the single-institution dataset of only 118 patients, all contoured by a single radiation oncologist. Larger, multi-center datasets with multiple expert contours as ground truth are needed to validate generalizability. The researchers also note that improved GPU hardware would further reduce processing time and potentially improve performance.
The study demonstrates that 3D nnU-Net can successfully automate the contouring of all seven relevant pelvic structures for prostate radiotherapy planning on CT images, with clinically acceptable dosimetric accuracy for most structures.
By reducing segmentation time from 20-30 minutes to under 2 minutes, this approach has the potential to significantly reduce radiation oncologist workload, improve consistency across patients and institutions, and accelerate the overall radiotherapy planning workflow.
The authors envision this tool as a supportive system for radiation oncologists rather than a replacement -- clinicians would review and approve the automatically generated contours before proceeding to treatment planning, maintaining clinical oversight while benefiting from AI efficiency.