Experiences with auto-contouring

[TFI1985]

Hi everyone!

I was wondering if you could share your experiences with the different available auto-contouring solutions on the market. I'd be very interested to know what software you are using, and if you think it facilitates and/or accelerates your work?

Specifically, the clinic where I work at is currently looking at solutions offered by Elekta (ABAS), Varian (Smart Segmentation), Brainlab (Elements) and possibly MIM Maestro, all with their various features and additional modules for both contouring and planning (e.g. Varian has a built in atlas for reference, which I imagine could be quite useful for residents such as myself?). I talked to a few colleagues at other clinics and the response is quite mixed, so I thought maybe I could get some additional feedback here

Thanks in advance!

---

Members don't see this ad.

 

[scarbrtj]

Mixed to negative experiences, to the point I am not impressed and am certainly not a customer or purchaser of autocontour solutions. I have said that autocontour is as frustrating as autocorrect (on the iPhone)--does what you want sometimes, but other times not so much. The *entire* point of contouring is to decide where dose should or shouldn't go (contouring is *not* for the purposes of defining anatomy--leave that to Netter). This seems to me to be the fundamental job description of a radiation oncologist and is thus a task I'm loathe to pawn off on a computer lest I get complacent or a false sense of security.

 

[seper]

we need a major breakthrough in algorithms research for this to become usable

 

[napoleondynamite]

I find the MIM co-pilot tool to be quite useful, particularly for H&N volumes. It's a real time saver for me. Not truly auto-contouring, but probably saves me an hour or two in some of the more complicated plans I do.

 

[thesauce]

Takes me longer to fix the contours than to do them from scratch

 

[TFI1985]

Thanks for your replies! They are sort of in touch with what I heard from colleagues at other clinics. I'd be interested to try the MIM Co-Pilot, will have to look into that.

Any other feedback is still greatly appreciated!

 

[cdf95cro]

We use ABAS/Admire, Velocity, and SPICE (Pinnacle) in our section (our prostate guys really like MIMVista, but I only know head and neck performance).

Based on or benchmarking data (Mohamed et al.) ABAS outperformed Velocity for everything except bone; soft tissue structure variation was significant. However, practically, ABAS/Admire lacks a GUI to view registrations, so we don't use it 2/2 clinical workflow constraints.

Clinically, we use Velocity or SPICE; we benchmarked the SPICE algorithm (Walker et al., attached, and found substantive time savings for residents; for some structures (e.g. brainstem, cord) its about as reliable as a human. For others (chiasm, cochleas) its worthless, so you still have to check.

My 2 cents:
-Autoseg software are time saving devices that should be carefully QA'd using your individual practice, and must be QA'd as a process *for each organ site*.
-Most autoseg programs are reasonably reliable for dose monitoring (e.g. tracking stuff you don't care enough about to contour accurately) or research applications as they have the advantage that if they are "wrong" regarding a ROI they are systematically wrong (as opposed to humans, who are both systematically and randomly "wrong" about segmentations/delineations) but if it is a critical ROI near tumor, you are better off contouring the OAR manually on any given individual case.
-Finally, most AS software are *terrible* for contouring de novo tumor/targets, so never use them for initial GTV/CTV/PTV segmentation unless you want to kill someone; they are only useful for registering intra-patient tumor changes on equivalent imaging w short intervals (e.g. adaptive RT), but even that can get tricky.

 

[radiaterMike]

Thanks for posting this. I suggest removing the PDFs and linking to pubmed since those articles are owned by the journal and should not be openly distributed (unless open access).

 

[Chartreuse Wombat]

Thanks for posting this. I suggest removing the PDFs and linking to pubmed since those articles are owned by the journal and should not be openly distributed (unless open access).

Ditto that

 

[TFI1985]

We use ABAS/Admire, Velocity, and SPICE (Pinnacle) in our section (our prostate guys really like MIMVista, but I only know head and neck performance).

Based on or benchmarking data (Mohamed et al.) ABAS outperformed Velocity for everything except bone; soft tissue structure variation was significant. However, practically, ABAS/Admire lacks a GUI to view registrations, so we don't use it 2/2 clinical workflow constraints.

Clinically, we use Velocity or SPICE; we benchmarked the SPICE algorithm (Walker et al., attached, and found substantive time savings for residents; for some structures (e.g. brainstem, cord) its about as reliable as a human. For others (chiasm, cochleas) its worthless, so you still have to check.

My 2 cents:
-Autoseg software are time saving devices that should be carefully QA'd using your individual practice, and must be QA'd as a process *for each organ site*.
-Most autoseg programs are reasonably reliable for dose monitoring (e.g. tracking stuff you don't care enough about to contour accurately) or research applications as they have the advantage that if they are "wrong" regarding a ROI they are systematically wrong (as opposed to humans, who are both systematically and randomly "wrong" about segmentations/delineations) but if it is a critical ROI near tumor, you are better off contouring the OAR manually on any given individual case.
-Finally, most AS software are *terrible* for contouring de novo tumor/targets, so never use them for initial GTV/CTV/PTV segmentation unless you want to kill someone; they are only useful for registering intra-patient tumor changes on equivalent imaging w short intervals (e.g. adaptive RT), but even that can get tricky.

Thank you!!