Computer-aided diagnosis...

[SaltySqueegee]

---

Members don't see this ad.

How far are we from automated computer screening of slides to make diagnoses?

Years, decades, beyond?
Quality control issues with the current systems?

Can we train a computer to view and interpret slides the same way a human pathologist perceives the slides?

-Salty

 

[Gene_]

Pap smears are already being screened by imagers. It takes a human to make a diagnosis, though. Pathology is just as much of an art as it is a science, so I don't see computers making a diagnosis in my lifetime.

 

[yaah]

I would say decades. The thing I find interesting though is that whatever capabilities the computer is going to have, it is going to be based on whoever is training it. What if the trainer has a low threshold for dysplasia? What if the trainer gets a diagnosis wrong? The first likely use for computer aided diagnosis will probably be in the differential diagnosis realm, or in cytology for identifying potentially malignant cells faster than a screener. I would probably base things on the radiology experience - it is going to be a lot easier to get computers to diagnose radiology studies than for path slides - and we are not there yet by a long shot. I mean, I would think that someone would have been able to teach a computer to rapidly interpret a chest xray by now, right?

 

[djmd]

I would say decades. The thing I find interesting though is that whatever capabilities the computer is going to have, it is going to be based on whoever is training it. What if the trainer has a low threshold for dysplasia? What if the trainer gets a diagnosis wrong? The first likely use for computer aided diagnosis will probably be in the differential diagnosis realm, or in cytology for identifying potentially malignant cells faster than a screener. I would probably base things on the radiology experience - it is going to be a lot easier to get computers to diagnose radiology studies than for path slides - and we are not there yet by a long shot. I mean, I would think that someone would have been able to teach a computer to rapidly interpret a chest xray by now, right?

Reminds me of a story about a computer "AI" trained to spot camouflaged tanks... They trained it on pictures of trees and bushes and trees with camo tanks... Till it was perfect...
then they tested it and it failed utterly.

Then they noticed all the tank pictures were sunny and the tree pictures were cloudy....

(sounds false, but I swear it was true, I can't source it because it dates to something I saw in undergrad days)

 

[SaltySqueegee]

Reminds me of a story about a computer "AI" trained to spot camouflaged tanks... They trained it on pictures of trees and bushes and trees with camo tanks... Till it was perfect...
then they tested it and it failed utterly.

Then they noticed all the tank pictures were sunny and the tree pictures were cloudy....

(sounds false, but I swear it was true, I can't source it because it dates to something I saw in undergrad days)

I recall a documentary on Discovery or TLC some time back where they follow where a person looks at on a screen to see where their eyes track for marketing purposes. E.g. Flash a picture of a model up and the guy automatically tracks his eyes to the appropriate physical parts he's intrinsically drawn to. A way to investigate good marketing images for commercials.

What if they flashed hist/cyto images on a screen to see what a pathologists eyes tracked towards?

-Salty

 

[SaltySqueegee]

Came across this interesting article on pubmed that describes eye tracking with pathologist attendings, residents and medical students with breast biopsies.

Eye-movement study and human performance using telepathology virtual slides: implications for medical education and differences with experience.

Krupinski EA, Tillack AA, Richter L, Henderson JT, Bhattacharyya AK, Scott KM, Graham AR, Descour MR, Davis JR, Weinstein RS.
Department of Radiology, University of Arizona College of Medicine, Tucson, AZ 85724, USA.

Hum Pathol. 2006 Dec;37(12):1543-56.

 

[djmd]

What if they flashed hist/cyto images on a screen to see what a pathologists eyes tracked towards?

The color purple....