Radiation therapy (RT) is a major treatment for prostate cancer, which is one of the most common cancers in men. Planning a course of radiation involves two critical steps: first, identifying and outlining the tumor and surrounding organs on a CT scan, then designing a radiation delivery plan that maximizes dose to the cancer while protecting healthy tissue.
The first step, called segmentation or contouring, requires a radiation oncologist to manually draw the boundaries of the tumor target (called the clinical target volume or CTV) and organs at risk such as the bladder, rectum, and femoral heads. This is time-consuming and subject to significant variation between different clinicians.
Inter-observer variability in prostate contouring is a well-documented problem. Different physicians draw slightly different boundaries around the same prostate on the same CT scan, which can translate into different radiation doses being delivered to different patients even when the intention is identical. This inconsistency undermines the reliability and fairness of treatment.
Machine learning and deep learning algorithms have shown promise for automating segmentation, potentially delivering more consistent contours and freeing up clinical time. However, most prior work automated either contouring or plan optimization separately. This study tested a fully automated workflow that handles both steps from a single CT scan with no manual input required.
The automated workflow was built within the RayStation treatment planning system. It uses a deep learning segmentation tool called RLS Male Pelvic to automatically identify the prostate, bladder, rectum, and femoral heads on a CT scan. A machine learning planning model called RSL-Prostate-6000 then predicts the ideal radiation dose distribution and generates a deliverable treatment plan.
The entire process is initiated with a single mouse click after importing the patient's CT scan. From start to finish, the workflow completes in approximately 10 minutes without any human intervention, in contrast to traditional manual planning that can take considerably longer and involves multiple clinical staff.
The study enrolled five patients with low-risk prostate cancer who had already been treated with 76.5 Gy of radiation delivered in 34 fractions. For each patient, six experienced radiation oncologists independently drew manual contours, providing a range that represents typical human variation in clinical practice.
Each combination of contour set and treatment plan was evaluated using standard dosimetric metrics including Dice Similarity Coefficient (DSC) and Hausdorff Distance for contour overlap, plus coverage indices like V95, D95, conformity index, and homogeneity index to assess dose delivery quality.
The AI-generated CTVs were significantly smaller than those drawn by human experts, with a median volume of 47.1 cm3 compared to 62.7 cm3 for the manual contours. The resulting planning target volumes (PTVs) were similarly smaller (105.7 vs. 138.4 cm3).
Despite the size difference, expert review found both the manual and AI-generated contours to be anatomically correct. No AI contour was identified as anatomically wrong. The most visible differences occurred at the boundary between the prostate and the seminal vesicle bases, as well as at the prostate apex, areas that are notoriously difficult to delineate on CT imaging.
Quantitative overlap analysis showed a Dice Similarity Coefficient of 0.86 for AI versus expert contours, compared to 0.90 for expert versus expert comparisons. While this difference was statistically significant, it falls within ranges reported in the broader literature for deep learning segmentation tools. The average Hausdorff Distance for AI contours was 1.27 mm, slightly above the 0.98 mm seen between human experts.
Importantly, the AI system produced highly consistent results each time it ran on the same images. When the segmentation was repeated, the AI generated identical contours, which is a key advantage over human drawing, which tends to vary even when the same physician reviews the same scan twice.
The automated treatment plans produced dose coverage of the target volumes that fell within the 95 percent confidence interval of the variation seen among manually created plans. In other words, the AI plans were no more different from the average expert plan than one expert plan was from another.
However, the AI plans showed meaningfully lower coverage of the manually drawn target volumes in the PTV. The deviation in D95 for the PTV was -14.77 percent for AI plans versus -6.49 percent for manual plans, largely because the AI contoured a smaller CTV and therefore designed a plan to cover a smaller area.
A notable advantage of the AI plans was significantly better sparing of organs at risk. The AI plans delivered less radiation to the bladder and rectum: the volume of bladder receiving at least 50 Gy was 9.42 percent on average for AI plans compared to 14.8 percent for manual plans, and similar reductions were seen for the rectum.
This organ-sparing benefit is likely a direct consequence of the smaller CTV used by the AI. With a smaller target to cover, the optimizer can design plans that more aggressively curve radiation away from adjacent organs, potentially reducing the risk of long-term treatment side effects like bladder irritation or rectal bleeding.
The most consequential finding of this study is that the AI systematically drew smaller target volumes than experienced clinicians. This is not necessarily wrong, since no AI contour was deemed anatomically incorrect, but it reflects real differences in how the AI versus human experts interpret ambiguous boundaries on CT images.
The AI segmentation tool was not trained on data from this specific institution. Inter-institutional differences in contouring guidelines, local anatomy conventions, and training cases likely contributed to the systematic volume discrepancy. Retraining the AI on local data would be expected to improve agreement.
The study highlights that the problem of inter-observer variability is not solved by AI, but transformed. Instead of variation between human physicians, there is variation between the AI system and any given human physician. The key question becomes which contour is more clinically appropriate, a question this study cannot definitively answer.
An important limitation is that only five patients were studied, all with typical anatomy and no unusual features like hip prostheses. The workflow has not been tested in challenging anatomical situations where automated tools are more likely to fail, meaning expert review remains essential for now.
The most immediate clinical benefit of this workflow is speed. Generating a full radiation treatment plan in approximately 10 minutes with no human involvement compares favorably to the hours typically required for manual planning, which involves a radiation oncologist contouring structures and a radiation therapist optimizing the plan.
Consistency is the other major benefit. Because the AI applies the same algorithm to every scan, it eliminates the day-to-day variation that occurs when different team members handle different patients. This could help institutions maintain quality standards across a large volume of cases.
All plans generated in this study were reviewed and approved by a board-certified radiation oncologist before being used for evaluation, and this expert review step would remain essential in clinical deployment. The AI workflow is therefore best described as a decision support tool that reduces manual workload and provides a high-quality starting point, rather than a fully autonomous replacement for clinical judgment.
Future work should include larger patient cohorts, patients with more complex anatomy, and sites with hip implants or unusual pelvic features. Training the segmentation model on institution-specific data would likely close the gap in target volume sizes seen in this study.
This study demonstrates that a fully automated, one-click machine learning workflow can generate radiation treatment plans for prostate cancer that are clinically acceptable and fall within the range of natural variation between human expert plans.
The AI workflow successfully integrated two components that had previously been automated only separately: target volume delineation and treatment plan optimization. Linking these into a seamless pipeline is a meaningful advance toward the vision of truly automated radiotherapy planning.
The finding that AI contours are systematically smaller than expert contours is an important area for improvement. Addressing this through institution-specific training data and better handling of ambiguous anatomical boundaries will be critical before this workflow can be used without close expert oversight.
Despite these limitations, the results support continued development and prospective testing of automated planning workflows. As these tools mature, they have the potential to free up significant clinical time, reduce treatment variability, and ultimately deliver more consistent and equitable care to prostate cancer patients.