The setting
Mayo Clinic Arizona’s prostate planning protocol uses anatomy that almost no vendor’s auto-segmentation model has seen during training: iodinated rectal spacers and rectal balloons. The shape and contrast of male pelvis structures changes substantially.
Three major vendors’ built-in deep-learning auto-segmentation models were evaluated. All three underperformed on the Mayo protocol — particularly on prostate, seminal vesicles, bladder, rectum, and femur heads.
The intervention
Rather than tune around vendor limitations, Mayo’s team used INTContour’s incremental retraining capability. They trained a new model on 100 locally-acquired cases that included the spacer and balloon configurations.
INTContour’s incremental learning runs entirely within the institution. No data leaves. No code is written — the training pipeline is part of the platform.
The readout
Mayo conducted a blinded 5-point clinical-acceptability rating on 115 test cases, scored by 6 GU radiation oncologists plus 2 RO residents.
The headline finding:
AI contours rated ≥ manual contours in ≥50% of cases — across every structure tested.
Per-structure detail (lowest to highest acceptance):
| Structure | AI ≥ manual rating |
|---|---|
| Seminal vesicles | 70% |
| Penile bulb | 73% |
| Prostate | 78% |
| Rectum | 82% |
| Femur heads | 88% |
| Bladder | 95% |
The lowest acceptance rate (seminal vesicles in junior-resident review, 63.8%) still clears the 50% threshold. The highest (bladder, RO1 review) is 100%.
The point
This is what “research today, FDA-cleared product tomorrow” looks like in practice. Mayo retrained the same FDA-cleared INTContour engine on the protocol their clinic actually uses — without writing code, without exporting data, without procuring a new vendor.
It’s also the strongest evidence we have for the platform claim: when the vendor model doesn’t fit the local protocol, incremental retraining inside the same platform closes the gap.
Reference
Follow-up study published as Duan et al., Radiotherapy & Oncology, 2025 — CT-only prostate auto-segmentation without MR.