Very intetesting TED talk about computer learning, some incredible advances in the past few years, with specific references to pathology- what are your thoughts? Pathologists replaced by computers in the next few generations?
Computer learning - the end of pathology as we know it?
- Thread starter MedUK
- Start date
Not a chance.
Image analysis in pathology is way too complex for a computer. Psychiatrists will be replaced by computers way before pathologists. Also, pattern recognition and image interpretation is just a part of pathology. You have to order and interpret ancillary studies, troubleshoot technical problems, know how to intelligently say that there's a grey area in a report, etc.
However computers can help pathologists, especially when it comes to cytopath screening. In a few decades computers will be another of our diagnostic tools, nothing more.
The video is too long to watch, at what point does Garfunkel talk about pathology?
Diagnostic pathology will be incredibly difficult to "replace" by "artificially intelligent" computer systems and its potential replacement is usually the fantasy of oncologists or surgeons that want to circumvent a true diagnosis and move directly to use poisons (oncotherapeutics) or bombs (surgeons). Unfortunately pathology is so poorly taught in medical school that med student go into specialties and have very little understanding of what a surgical pathologist does. They see a generalized image of PTC in Robbins and think that diagnostic surgical pathology is straight forward and easy. So many medical students, residents, and doctors have absolutely no idea what we do so they listen to talks like this and feel like its a threat to our field. It is possible that some technologists may be replaced or assisted by these upcoming trends..... but if machines make diagnostic and prognostic actionable calls that lead to bombs and poisons without a physician human interface we are all screwed.
- Joined
- Aug 10, 2006
- Messages
- 410
- Reaction score
- 27
TED talks aren't what they used to be. Nowadays I feel like a lot of them are just kitsch or weird or too fluffy.
Oh ye of little faith. Cars can drive themselves. Planes can fly themselves. In fact, they can do a better job than most humans. They don't because human pilots have a fear of not being in control. When was the last time human-machine chess matches were even in the news? Etc.
Queue the outrage of the inevitable....
Having self driving cars will be nice... it will be especially nice in Colorado so you can do bong hits behind the wheel and not deal with stoned driving laws. even if you train a computer to look at a biopsy, its unlikely that anyone would want to be poisoned or bombed without some sort of human interface. Maybe you are right and it will become like laboratory automation... the hardest part will be training a computer or machine to cut 3-5 micron sections. So we'll have histotechs working for the machines....Maybe pathologists will have to do the cutting. Unless there is some sort of spectral imaging technology that does away with paraffin embedding. Shoot... maybe there will be nanobots with lasers that can target tumors after being systemically introduced through the retro-orbital vascular plexus. Interestingly one of the purported technologies that was going to cause a shift in pathology and possible lead to outsourcing (digital pathology)is becoming less popular for surgical pathologists. Polls show that although groups use it for education, routine sign out of SP is cumbersome than slapping glass.
Last edited:
- Joined
- Dec 16, 2010
- Messages
- 2,130
- Reaction score
- 1,032
There are 3D printers that can build houses now...
http://www.cnet.com/news/worlds-first-3d-printed-apartment-building-constructed-in-china/
If you would have told me that the field of cytopathology would be reduced to rubble in 2014, I wouldn't have believed you. Now most labs have about a 50 percent (or larger) drop in cytopath specimens in just a few years ago. All it takes is a few disruptive technologies to really harm your accession count and salary.
http://www.cnet.com/news/worlds-first-3d-printed-apartment-building-constructed-in-china/
If you would have told me that the field of cytopathology would be reduced to rubble in 2014, I wouldn't have believed you. Now most labs have about a 50 percent (or larger) drop in cytopath specimens in just a few years ago. All it takes is a few disruptive technologies to really harm your accession count and salary.
- Joined
- Mar 17, 2003
- Messages
- 2,962
- Reaction score
- 79
There are 3D printers that can build houses now...
http://www.cnet.com/news/worlds-first-3d-printed-apartment-building-constructed-in-china/
If you would have told me that the field of cytopathology would be reduced to rubble in 2014, I wouldn't have believed you. Now most labs have about a 50 percent (or larger) drop in cytopath specimens in just a few years ago. All it takes is a few disruptive technologies to really harm your accession count and salary.
But the point of what we do is to help people. It is not for us to have lots of accessions for the sake of billing.
If a more accurate or less invasive, less harmful way of helping patients is developed, what else can we do but accept it, even celebrate it.
Last edited:
- Joined
- Oct 3, 2014
- Messages
- 127
- Reaction score
- 113
I don't see computers replacing pathologists for diagnostics. I do however see them as becoming more and more important in helping us do our job more quickly,efficiently, and producing more meaningful data than we do now. In effect, I see computer aided pathology as what is going to happen rather than computer performed pathology. For example -- at molecular right now - there is no way we could sign out next gen cases without help from the computer organizing and structuring the data in the meaningful way.
On the AP side of things I agree with thebouque that image analysis is extremely complicated. Our institution uses digital slides for remote assessment of UNOS kidney biopsies but that is about it other than med student teaching. To get to the point of just digitizing slide information for routine clinical use has issues for any place that has non-trivial volume. Places like UPMC is doing interesting stuff with remote viewing of images for remote consultation, but that is not a general AP solution. Multiple barriers existing from the engineering to the regulatory level (speed of scanning, security/archiving/etc of scanned images/FDA ant its view of scanners as medical devices, etc...). The HCI between the pathologist and virtual slide software is about as good as the interface between clinicians and their EMR's (i.e. craptastic) which is another important item to resolve. I believe in time these will be overcome, but not in the next 5-10 years.
The computational complexity of image analysis (feature detection, etc...) is large but I try not to discount Moore's "law" with respect to speed of computation and the ability to overcome NP hard problems in a reasonable time before the heat death of the universe. I think algorithmic and hardware advances will occur that may help screen or flag features of slides (once we can routinely digitize them) slides but will not be able to replace a human diagnostician. From a practical standpoint, the validation of such a test would be exceedingly complicated in today's regulatory environment. And that might be the easiest part of the whole thing.
Ultimately I see the pathologist working in conjunction with a computer to form a diagnosis. For example, viewing digital (or possibly even glass slides that are digitized on the fly) with a HUD like display giving ancillary or additional information about a given field of view (e.g. mitotic count, nuclear pleomorphism index, KI-67 index, etc...) and then using the computer to automatically record staging/grading pertinent information for transmission to the relevant target EMR (no paper reports in this model...) I do not see this type of streamlined use coming anytime soon. Maybe in 15-20 years or so. Either way, I do not see the pathologists being replaced anytime soon.
- chooks
On the AP side of things I agree with thebouque that image analysis is extremely complicated. Our institution uses digital slides for remote assessment of UNOS kidney biopsies but that is about it other than med student teaching. To get to the point of just digitizing slide information for routine clinical use has issues for any place that has non-trivial volume. Places like UPMC is doing interesting stuff with remote viewing of images for remote consultation, but that is not a general AP solution. Multiple barriers existing from the engineering to the regulatory level (speed of scanning, security/archiving/etc of scanned images/FDA ant its view of scanners as medical devices, etc...). The HCI between the pathologist and virtual slide software is about as good as the interface between clinicians and their EMR's (i.e. craptastic) which is another important item to resolve. I believe in time these will be overcome, but not in the next 5-10 years.
The computational complexity of image analysis (feature detection, etc...) is large but I try not to discount Moore's "law" with respect to speed of computation and the ability to overcome NP hard problems in a reasonable time before the heat death of the universe. I think algorithmic and hardware advances will occur that may help screen or flag features of slides (once we can routinely digitize them) slides but will not be able to replace a human diagnostician. From a practical standpoint, the validation of such a test would be exceedingly complicated in today's regulatory environment. And that might be the easiest part of the whole thing.
Ultimately I see the pathologist working in conjunction with a computer to form a diagnosis. For example, viewing digital (or possibly even glass slides that are digitized on the fly) with a HUD like display giving ancillary or additional information about a given field of view (e.g. mitotic count, nuclear pleomorphism index, KI-67 index, etc...) and then using the computer to automatically record staging/grading pertinent information for transmission to the relevant target EMR (no paper reports in this model...) I do not see this type of streamlined use coming anytime soon. Maybe in 15-20 years or so. Either way, I do not see the pathologists being replaced anytime soon.
- chooks
- Joined
- Aug 15, 2003
- Messages
- 28,059
- Reaction score
- 438
This may be down the horizon, but most likely this is going to impact radiology before it impacts pathology. Radiology images are simpler, in a sense, for computers to decipher. Also, this has been attempted (rather successfully) with pap smears but cytotechs have not gone away, nor has visual diagnosis of pap smears. So yes, the above comments are more likely - computer-aided diagnosis but not replacement. However, one can never predict the future. Theoretically morphology on biopsies could go away totally and be replaced by genetic analysis.
- Joined
- Oct 17, 2002
- Messages
- 362
- Reaction score
- 4
I'm not overly concerned with AI and algorithmic computer vision diagnosis in the short-to-medium term future. I have no doubts that eventually anatomic pathology diagnostics as we currently practice will be eliminated by either advances in computer vision/AI or some as-currently undeveloped technological advances (multispectrum in-situ radiology techniques or something). All of us currently practicing or in the pipeline for training will likely be retired / near-retired by the time of widespread clinical adoption.
Barriers to AI/algorithmic adoption of tasks generally include:
1. Creativity tasks
2. Social Intelligence tasks
3. Manipulation of complex objects/perception tasks
Diagnostic pathology is probably well insulated against at least 2, if not all, of those 3 barriers.
I think a much more immediate threat to our employment as pathologists will be the development of two things: 1. A national medical license allowing easy multi-state practice, and 2. FDA approval of primary diagnosis through digital pathology.
My gut feeling is that the current national workload of diagnostic pathology could be easily handled by as few as 40-50% of currently practicing pathologists if you created a high efficiency pipeline of specimen acquistion -> grossing/processing/histo -> scanned -> interpreted by remote pathologists.
This would be a prime target for bean-counting under a national socialized/single-payor medical delivery system.
Barriers to AI/algorithmic adoption of tasks generally include:
1. Creativity tasks
2. Social Intelligence tasks
3. Manipulation of complex objects/perception tasks
Diagnostic pathology is probably well insulated against at least 2, if not all, of those 3 barriers.
I think a much more immediate threat to our employment as pathologists will be the development of two things: 1. A national medical license allowing easy multi-state practice, and 2. FDA approval of primary diagnosis through digital pathology.
My gut feeling is that the current national workload of diagnostic pathology could be easily handled by as few as 40-50% of currently practicing pathologists if you created a high efficiency pipeline of specimen acquistion -> grossing/processing/histo -> scanned -> interpreted by remote pathologists.
This would be a prime target for bean-counting under a national socialized/single-payor medical delivery system.
