Before a prostate cancer patient can begin radiation therapy, the treatment team must carefully draw boundaries on CT scan images to define two types of structures: the Clinical Target Volume (CTV) -- the prostate and seminal vesicles where cancer must be treated -- and the Organs at Risk (OARs) -- nearby healthy structures like the bladder, rectum, and femoral heads that must be protected from excessive radiation dose.
This process, called contouring or segmentation, is performed manually by radiation oncologists and can take 15 to 25 minutes per patient. It is not just time-consuming but also inherently variable: different clinicians frequently draw slightly different boundaries around the same structures, even when following the same guidelines. This inter-observer variability introduces uncertainty into treatment planning that can affect both how well the cancer is treated and how much normal tissue is irradiated.
AI-based automatic contouring tools have been developed over the past decade with the goal of reducing both the time burden and the variability. The challenge is validating these systems rigorously in real clinical settings -- not just in controlled retrospective comparisons -- to confirm they work reliably with the messy, variable data of everyday practice.
This study from the IRCCS San Raffaele Scientific Institute in Milan evaluated a commercially available AI auto-contouring system (MIM Protege, v1.1.2, using a U-Net model trained on multi-institutional data) in a real clinical workflow. Twenty intermediate- and high-risk prostate cancer patients were selected from those treated at the institution within the previous year with standard CT-based planning.
Two radiation oncologists, each with more than 10 years of prostate contouring experience, independently drew manual contours for five structures on each patient's planning CT: bladder, rectum, femoral heads including femurs, prostate, and seminal vesicles. The same CT images were also sent through the AI system to generate automatic contours, which the oncologists could then edit as needed -- mirroring the intended real-world workflow.
The study evaluated three scenarios: automatic contours without any editing, automatic contours after manual editing by each observer, and purely manual contours. Performance was measured using the Dice Similarity Coefficient (DSC) -- a score from 0 to 1 measuring overlap between two contours -- and the Hausdorff Distance (HD), which measures the maximum boundary discrepancy. Oncologists also rated the clinical quality of AI contours on a 1 to 5 scale, and contouring times were recorded precisely.
For unedited AI contours, Dice scores ranged from 0.65 for seminal vesicles to 0.94 for bladder, with prostate at 0.82, rectum at 0.82, and femoral heads at approximately 0.82 on average. These values are at or within the range of normal inter-observer variability for manual contouring -- meaning the AI's automatic output was as accurate as a second independent human reader would be for most structures.
Seminal vesicles were the most challenging structure for the AI, which is expected -- they are small, variable in shape, and closely adjacent to other structures. Bladder and femoral heads performed best because these structures have clear, consistent boundaries on CT that are relatively easy for any observer (human or AI) to identify.
The clinical quality scores confirmed the quantitative results. The median score was 4 (minor editing required) for both observers. Only 2-3 out of 20 patients per observer received a score of 3 or lower, meaning substantial editing was rarely needed. No patient required completely discarding the AI contour and starting from scratch, suggesting the AI output provided a useful starting point even for the cases where it was least accurate.
After the radiation oncologists edited the AI-generated contours, performance improved significantly across all structures. The largest gains were seen for seminal vesicles (DSC improved from 0.65 to 0.76) and the prostate (from 0.82 to approximately 0.86). These improvements were statistically significant (p less than or equal to 0.002), confirming that even brief targeted editing meaningfully improves contour quality beyond what the AI produces alone.
A critical finding was that after editing, AI-assisted contours were more consistent between the two observers than their purely manual contours. For seminal vesicles, the inter-observer DSC improved from 0.73 (manual) to 0.83 (AI-assisted and edited). For prostate, it improved from 0.83 to 0.88. For rectum and bladder, improvements were even more striking. This means the AI acted as a standardizing anchor, guiding both observers toward more similar final contours than they would have produced working entirely independently.
The explanation for this reduction in inter-observer variability is that when two oncologists start from the same AI-generated template and each makes small adjustments to correct it, their final contours tend to be more similar than if they had independently drawn contours from scratch. The AI effectively introduces a common reference that reduces divergence from purely personal interpretation styles.
The time savings from AI-assisted contouring were substantial and statistically significant (p less than 0.01) for both observers. Observer 1 reduced contouring time from an average of 24 minutes for manual contouring to 7 minutes for AI editing -- a saving of about 17 minutes per patient. Observer 2 went from 17 minutes down to 3 minutes, saving 14 minutes per patient.
The 8-minute automated AI processing time is not counted as operator time because the segmentation runs in the background without any human attention. With the fully automated clinical workflow now implemented at San Raffaele, the entire process from CT upload to completed contours requires less than 8 minutes total, with oncologist interaction time under 2 minutes for most patients.
This time saving is particularly meaningful in high-volume radiotherapy departments where planning CT scans may be processed for 10 or more patients daily. Saving 15 minutes per patient translates to hours of freed oncologist time each day -- capacity that could be redirected toward more complex patient cases, teaching, or clinical research.
The study demonstrates that AI auto-contouring can replace manual contouring in routine prostate cancer radiotherapy planning, with the important caveat that a radiation oncologist must review and potentially edit the automated contours before approving them for treatment. This human-in-the-loop model -- AI generates, human verifies -- represents the appropriate clinical framework at the current state of the technology.
The San Raffaele group has now fully integrated this AI contouring system into their clinical workflow. Once a patient's CT is imported into the treatment planning system, it is automatically sent to the auto-contouring workstation, and the resulting contours are returned within 8 minutes for oncologist review. The institution's goal is to combine this automated contouring pipeline with their existing automated treatment planning system to create a nearly fully automated treatment preparation chain.
Comparison with other published studies shows that the results obtained here are generally consistent with or slightly better than previously reported for similar systems. The reduction of inter-observer variability documented in this study -- rarely reported in prior literature -- is an important new contribution demonstrating a benefit of AI contouring beyond simple time savings: more consistent treatment plans across different clinical providers.
The study is limited by its small size (20 patients, two observers, one institution). While the results are clear and consistent, broader multi-institutional validation would strengthen confidence in the findings and help determine whether performance is reproducible with different AI systems, different CT scanner protocols, and different clinical practice patterns across institutions.
The current AI model was validated only for the five structures routinely contoured in localized prostate cancer treatment. Extending the workflow to patients who need pelvic lymph node irradiation -- a more complex treatment requiring contouring of additional nodal regions and bowel loops -- will require training the AI on additional anatomical targets not yet covered by the validated system.
Despite these limitations, the study provides compelling evidence that AI auto-contouring improves every measurable aspect of the prostate cancer radiotherapy planning workflow: contour accuracy matches human performance, editing improves it further, inter-observer consistency improves significantly, and oncologist time is reduced by roughly 70%. This combination of benefits supports adoption as part of the standard clinical treatment preparation process.