Contouring on MRI for Prostate
Cancer Treatment
CT scans: the old way
For decades, radiation oncologists have traditionally used CT scans to design treatment planning. But CT scans do not show the precise boundaries of the prostate. One study showed that oncologists using CT scans tend to draw the “prostate” 30% too big while still only covering 84% of the actual organ.
Does high accuracy matter?
Well, there is no benefit to treating the wrong spot, and it might be bad to miss part of the prostate. Also, the MIRAGE trial showed that a 2 mm change in planning margin increases side effects, so small changes can potentially impact patient outcomes.
MRI: improved imaging
MRI is much better for seeing the prostate. MRI is also necessary if you want to protect the urethra and if you want to deliver a targeted focal boost of radiation dose to the tumor. The FLAME trial showed that targeted focal boost improves cure rates for prostate cancer.
Need for physician education
Many oncologists are not as familiar with MRI as they are with CT scans. We decided to define a reference standard for prostate MRI. The goals were (1) to create a contouring atlas for physician education and (2) to create a dataset with expert contours that could be used to evaluate the accuracy of artificial intelligence (AI) tools. This work is described in our study here.
How we defined the reference standard
A panel of four experienced prostate cancer specialists created the reference standard: Dr. Tyler Seibert (radiation oncologist), Dr. Robert Dess (radiation oncologist), Dr. Eric Weinberg (radiologist), and Dr. Daniel Margolis (radiologist). One of the radiation oncologists (Seibert) meticulously outlined the prostate on high-resolution axial T2-weighted MRI. The panel then met to review each slice at least twice. All three planes were always reviewed, and high-resolution coronal and sagittal images were available as needed. If any panel member saw the contours deviate from the true prostate boundary by more than a millimeter, the panel discussed and deliberated until a unanimous agreement was reached.
Some technical details / ground rules
Before the consensus review, the panel decided on ground rules for situations with more than one reasonable approach. If partial volume effects were present but the prostate was clearly visible, this was included as prostate. The proximal seminal vesicles (SVs) were not intentionally included. If a visible plane (on axial images) divided the SVs from the prostate, the tissue was labeled as SV and excluded. Tissue with SV appearance but without a clear dividing plane on an axial slice was included as prostate. While this may lead to inclusion of a small proportion of SVs, a rule was needed, and over-estimation was preferable to undercovering the prostate. Additionally, the panel agreed this tissue would be indistinguishable from prostate on CT and was invariably included in prior trials of prostate cancer radiotherapy utilizing CT for treatment planning. The tip of the apex may not be perfectly visualized on MRI as (per PI-RADS v2.1) axial slices are recommended to be 3.0 mm thick (in-plane resolution should be 0.7×0.4 mm). Therefore, at the apex, any suspicion of visible prostate, even with only subtle partial volume effects, was included as prostate.
What about the urethra?
The urethra is challenging to reliably contour on prostate MRI. We defined a reference standard for the urethra using the same panel review as for the prostate. The panel did not have the same level of certainty about the precise boundaries of the urethra as it did for the prostate, but consensus was achieved for the trajectory through each prostate.
Contouring atlases
A few full example cases are available below. If you use these in presentations, trial protocols, etc., please cite the following papers.
Prostate atlases: Song, Y. et al. Multidisciplinary consensus prostate contours on magnetic resonance imaging: educational atlas and reference standard for artificial intelligence benchmarking. International Journal of Radiation Oncology*Biology*Physics (2025) doi:10.1016/j.ijrobp.2025.03.024.
Urethra atlases: Song, Y. et al. Urethra contours on MRI: Multidisciplinary consensus educational atlas and reference standard for artificial intelligence benchmarking. Radiotherapy and Oncology 111231 (2026) doi:10.1016/j.radonc.2025.111231.
​
Prostate Case 1: Download the PPT or PDF
Prostate Case 2: Download the PPT or PDF
Urethra Case 1: Download the PPT or PDF
​
What about AI tools?
We tested 6 different tools for detecting the prostate automatically. Overall, they are pretty good. The best of them (developed in our lab) is similar to human physician performance. See the full results here. For the urethra, we developed an AI model that outperformed human physicians. See the full results here and here. AI models can make significant errors, though, and their outputs should be carefully reviewed prior to using to treat a patient.