[Unknown Member]

None.

[afazio.uk_887]

They are all worthless currently.  

[satyanar]

Using? Zero.

Imbedded in the workflow? I see the AI hemorrhage detection software is on our non-con stroke CTs. Never seen it make a difference. Have seen it over call.

[abananarest]

None.

[Unknown Member]

Lung nodule cad or iCAD with neural network driven algorithm. Its pretty good for catching little nodules but will miss a 5cm perihilar mass (guessing it thinks its just part of mediastinum since not surrounded by king).

It basically is only useful when it finds zero-3 nodules cause you can speed up the nodule search a little bit. For me it takes away the MIP review time, I fast scroll through 1mms then either MIPS or the lung cad findings

[conormurray33]

Surprised at your answer. Try AI-doc or AI-rad companion (I have used these extensively) – I’d be surprised if you don’t find they help.

[tdetlie_105]

AIDoc…Pretty good for ICH, PE, free air, pneumtx on xray, displaced rib fxs…PE is the best thus far, if neg for PE my time spent looking is like 80% less than normal…also it acts as a triage as these cases will come up as potential positives so at the very least, you have a chance to look at these studies sooner than later. 

[skysdad]

I second AI doc. 

[amotter]

RadAI its an impression generator.

[btomba_77]

We have incorporated AZMed’s AI fracture detection tool.
 
For the first time it was actually helpful today.
 
I was reading a clavicle and getting ready to call it normal, but the AI tool saw a posterior 3rb rib fracture I would have missed.
 
 
 
All other cases have been either concordant with what I was going to say or false positive 

[22002469]

How long have you been using it? Does it never miss a fracture (false neg)?
 
At minimum, would be nice for triage when X-Rays get backed up. Can be a little awkward to read unsuspected acute fractures multiple days after the fact. 

[btomba_77]

We’re the main trial site. Been live on PACSs for a couple of months now.   It was my eyeballs (and those of the other MSK rads) that trained up the tool. 
 
So far no false negs that i’ve seen … but it calls a lot of false pos.
 
 

[22002469]

Gotcha.

Last question how often do you get false positives? Are we talking 5%? 10%? 50%?

[btomba_77]

Quote from Radsoxfan

Gotcha.

Last question how often do you get false positives? Are we talking 5%? 10%? 50%?

For sure >10% of the calls are false pos. …. old fractures, big osteophytes …
 
I’m not keeping the data but i’d guesstimate 20-ish %

[22002469]

Thanks. So if Im understanding correctly.

To pick some random numbers, if you had 1000 unread MSK X-rays on the list, this would flag roughly 200 as possible fractures that could be looked at preferentially sooner by a rad. The other 800 would be VERY unlikely to have any acute fracture and could be read later? Obviously ideally backups dont get that bad but sometimes what can ya do.

Of course, fractures are not the only potentially urgent finding on an MSK plain film. Not that X-rays are good for early osteo anyway, but if there is something to see does it flag that as well? (Guess previously it was not my last question :))

[22002469]

I guess if the false positive rate is 20% it would flag more than 200 in that example since it would also flag correctly the real acute fractures but similar idea.

[btomba_77]

Quote from Radsoxfan

Thanks. So if Im understanding correctly.

To pick some random numbers, if you had 1000 unread MSK X-rays on the list, this would flag roughly 200 as possible fractures that could be looked at preferentially sooner by a rad. The other 800 would be VERY unlikely to have any acute fracture and could be read later? Obviously ideally backups dont get that bad but sometimes what can ya do.

Of course, fractures are not the only potentially urgent finding on an MSK plain film. Not that X-rays are good for early osteo anyway, but if there is something to see does it flag that as well? (Guess previously it was not my last question :))

Yeah … that’s how the tool is being used.
 
On the pacs list it gives a flag of “suspicious finding” or “no suspicious finding” …
 
if you got really backed up you could read the suspicious cases first.

[stlmchenry_510]

Maybe Google should have named their new AI program Lolita in memorial to their late friend, Jeffrey Epstein.

[stlmchenry_510]

All these companies are in bed together and insider trade.

[leonardo-campos2804]

viz ai. horrible

[1telerad]

rScriptor

[conormurray33]

Siemens AI-RAD companion. I do A LOT of lung cancer screening work, it saves my butt in 2% of cases and increases my speed with preserved accuracy by c 20%.

[pa30pilot]

I like rScriptor and RAD.AI impressions. Use both daily.