Delivering radiation precisely to a prostate tumor while protecting surrounding organs requires accurate boundaries to be drawn on CT scan images for both the tumor and nearby healthy structures. This process -- called contouring or segmentation -- must identify the prostate (the target), the bladder, and the rectum (the organs at risk), because radiation dose to these structures must be carefully limited to prevent side effects like incontinence and rectal bleeding.
Manual contouring by a radiation oncologist is time-consuming, often taking several minutes per patient, and is subject to significant variability between physicians. Different doctors draw the same organ boundaries differently, introducing inconsistency in treatment plans. In clinical settings where patients may need daily re-imaging and re-planning -- such as adaptive radiation therapy -- this bottleneck becomes clinically significant.
Automated segmentation tools have been developed to address this problem. Early methods used atlas-based approaches -- warping labeled reference images onto new patients. More recently, convolutional neural networks (CNNs) have replaced these older techniques and achieved state-of-the-art performance on segmentation benchmarks. The 3D U-Net architecture, built on the volumetric V-Net framework, is a leading approach that processes the full 3D CT volume and produces complete organ boundary predictions automatically.
Despite rapid advances, evaluation of these auto-segmentation tools has focused almost exclusively on geometric metrics -- how closely the predicted contour matches the manual contour in terms of spatial overlap and boundary distances. What is missing from most published studies is the clinically more relevant question: do plans optimized on automatic contours deliver acceptable radiation doses to the tumor and normal tissues? This study was designed to answer that question.
The study used CT images from 69 low-grade prostate cancer patients treated at LMU Munich. Patients with metal hip implants or fiducial markers -- both of which cause artifacts that degrade CT image quality in the prostate region -- were excluded, as these represent a known challenge for automated segmentation. The dataset was split into 47 training, 11 validation, and 11 test cases.
A 3D U-Net was trained separately for three structures: prostate, bladder, and rectum. The architecture used five encoder levels with 16 to 256 channels and included residual connections -- shortcut pathways that help gradients flow during training and improve segmentation of complex boundaries. Training used Dice loss, augmentation including 3D rotations, translations, elastic deformations, and zooming, and images were resampled to 1 cubic millimeter isotropic voxels before processing.
After training, the key methodological innovation was generating actual radiation treatment plans for all 11 test patients using both the manual (expert) contours and the 3D U-Net predicted contours, with identical optimization settings. Plans used Volumetric Modulated Arc Therapy (VMAT) -- a precise rotational delivery technique -- targeting 74 Gy to the prostate in 37 fractions. This design isolates the effect of contour quality on treatment plan quality.
To ensure that any differences in plan quality reflected contour differences rather than solution ambiguity in the optimization algorithm, the study validated that small perturbations to the manual segmentation produced plans that varied by less than 2% -- confirming the optimization was stable. This is an important methodological safeguard rarely implemented in comparable studies.
On the 11 test cases, the 3D U-Net achieved mean Dice Similarity Coefficients (DSC) of 0.87 for the prostate, 0.97 for the bladder, and 0.89 for the rectum. Average boundary distances (Hausdorff distances) were below 1.6 mm for prostate, 0.95 mm for bladder, and 1.4 mm for rectum -- indicating that contour boundaries were within approximately 1-2 millimeters of the expert-drawn reference in most cases.
Bladder segmentation was most accurate, reflecting its simple roughly spherical geometry and large size. The prostate was most challenging because it has poor soft tissue contrast on CT -- the prostate and surrounding tissue appear similar in CT Hounsfield values, making boundaries difficult to identify even for experienced human observers. The rectum showed intermediate difficulty, with one outlier case where a portion of the colon was incorrectly classified as rectum due to the ambiguous boundary between these structures.
Comparing with published state-of-the-art methods, the 3D U-Net performed competitively. For prostate, DSC 0.87 was slightly below two competing methods achieving 0.90, but comparable to another reporting 0.86. For bladder and rectum, the performance was either equivalent or superior to alternative approaches. The results confirm that the trained network meets the current standard for automated prostate region segmentation on CT.
Patients excluded from training due to metal implants or fiducial markers were tested separately. Performance on these artifact-corrupted cases was severely degraded -- prostate DSC fell to 0.60 with average Hausdorff distance exceeding 32 mm -- demonstrating that artifact handling requires dedicated training data. This is a practical limitation that any clinical deployment of the system would need to address.
Dose to the organs at risk was the most clinically critical question. For bladder dose metrics (V60/65/70 Gy), plans optimized on automatic contours agreed with expert-contour plans within plus or minus 5%. For rectum (V50/65/70 Gy), agreement was within plus or minus 2%. Neither organ showed statistically significant dose differences, and there was no consistent trend toward higher or lower doses to these sensitive structures -- confirming that automatic contours did not systematically overdose normal tissues.
For the treatment target (prostate and its 3 mm planning margin), D98% (dose received by 98% of the volume), D2%, and V95% showed agreement within approximately 3 Gy or 2% in 10 of 11 patients. The one exception was a patient (10% of the test set) where the predicted prostate contour was shifted toward the bladder -- an error reflected in a relatively low DSC of 0.82 that translated into a D98% discrepancy of 14.59 Gy for the clinical target volume, a clinically unacceptable deviation.
The average gamma pass rate -- a composite metric evaluating both dose magnitude and spatial accuracy -- was 85%, with a range of 71 to 94% across the 11 patients. Gamma analysis is stricter than DVH metrics and captures spatial dose discrepancies near the planning target volume boundary, where dose gradients are steep. The gamma criterion of 3%, 3 mm is a commonly used clinical standard; values below 80% typically indicate plans that require human review.
The conformity index -- measuring how precisely the prescribed dose is confined to the target -- averaged 0.85 for expert-contour plans and 0.78 for U-Net-based plans. The lower values for automatic contours reflect that plans optimized on predicted contours were evaluated against the expert contours, creating a systematic comparison disadvantage. This methodological factor slightly exaggerates the apparent conformity difference.
The most important scientific finding of this study is that geometric metrics -- the standard method for evaluating auto-segmentation algorithms -- do not reliably predict how well the resulting treatment plans will perform dosimetrically. The Pearson correlation between prostate DSC and gamma pass rate was only 0.67 (moderate positive correlation), and no statistically significant correlations were found between DSC and the clinically critical DVH parameters for organs at risk or target coverage.
The paper provides concrete examples that illustrate this disconnect. Two patients with identical prostate DSC of 0.85 had gamma pass rates of 93% and 74% respectively -- the second being marginal. Similarly, patients with DSC of 0.88 and 0.91 had gamma pass rates of 94% and 87%. These cases demonstrate that a contour can achieve excellent geometric overlap while producing a substantially different dose distribution in clinically relevant regions.
This mismatch arises because DSC measures overlap across the entire contour volume, but what matters for radiation therapy is accuracy specifically at the boundaries -- particularly near the interface with sensitive organs. A small systematic shift at the prostate-rectum boundary can substantially change the dose received by the rectum without significantly changing the overall DSC value. Hausdorff distance is a better proxy but still insufficient.
The implication is direct: the field cannot rely on DSC alone to certify that an auto-segmentation tool is clinically safe. Dosimetric evaluation -- actually computing what dose distributions result from the automatic contours -- should be part of the standard validation workflow for any AI segmentation system intended for use in radiation therapy treatment planning.
The study demonstrates that 3D U-Net segmentation of prostate, bladder, and rectum on CT images achieves clinically acceptable treatment plan quality in the majority of cases, with correct organ-at-risk dose constraints and adequate target coverage. This result supports the feasibility of using AI auto-segmentation as a starting point for treatment planning, where a physician reviews and corrects the automatic contours rather than drawing from scratch.
The single outlier patient -- where prostate contour misplacement caused a 14.59 Gy underdosage of the tumor -- underscores why human review of automatic contours remains essential before clinical use. Until automated quality assurance tools can reliably flag contour errors, physician oversight is a non-negotiable safety requirement.
The application of automated segmentation most clearly relevant to this technology is online adaptive radiation therapy -- workflows where patients are re-imaged and plans are re-optimized daily to account for bladder filling, rectal distension, and prostate movement. In this setting, recontouring speed is critical, and even a slightly imperfect automatic contour corrected in 1-2 minutes is far preferable to a manual contour requiring 10-15 minutes.
Future extensions of this work should include seminal vesicle segmentation to support higher-stage prostate cancer patients, dedicated training for artifact-corrupted images from patients with metal implants, and integration with dosimetric quality control metrics that can automatically flag cases where the plan deviates from clinical constraints -- enabling a human review workflow that prioritizes attention where it is most needed.