Slashdot Mirror
AI Can Diagnose Prostate Cancer As Well As a Pathologist (sciencebusiness.net)
An anonymous reader quotes a report from Science Business: Chinese researchers have developed an artificial intelligence system which can diagnose cancerous prostate samples as accurately as any pathologist, holding out the possibility of streamlining and eliminating variation in the process of cancer diagnosis. The system may also help overcome shortages of trained pathologists and in the longer term lead to automated or partially-automated prostate diagnosis. Confirmation of a prostate cancer diagnosis normally requires a biopsy sample to be examined by a pathologist. Now the Chinese AI system has shown similar levels of accuracy to pathologists and can also accurately classify the level of malignancy of the cancer, eliminating the variability which can creep into human diagnoses. [Hongqian Guo, who led the research group] took 918 prostate samples from 283 patients and ran these through the AI system, with the software gradually learning and improving diagnosis. The pathology images were subdivided into 40,000 smaller samples of which 30,000 were used to train the software while the remaining 10,000 were used to test accuracy. The results showed an accurate diagnosis in 99.38 per cent of cases, using a human pathologist as a gold standard. Guo said that means the AI system is as accurate as a pathologist. The research was presented at the 33rd European Association of Urology Congress in Copenhagen.
So you just point a camera at the person and they tell the AI the symptoms? Or do they mean you have a doctor get the symptoms and AI (google search) looks up matching diseases.
I'm not sure I'm ok with a robot examining me like that. I prefer the awkward wriggling fingers of a real human, and the incredibly strained communication before and after.
Some things do require a human touch.
Like my prostrate.
Velociraptor = Distiraptor / Timeraptor
I prefer a more natural approach to the prostate exam.
You are welcome on my lawn.
I can see this as a method to speed up processing samples and marking ones for evaluation
It's not a typo if you understood the meaning!
"As Well As a good Pathologist" would be more accurate.
Slashdot, fix the reply notifications... You won't get away with it...
Yes, but cars don't go up the butt. This will save millions of homophobes from being penetrated by another man.
not sure I like the sound of that
Table-ized A.I.
Can AI misdiagnose better than a pathologist?
AI Can Diagnose Prostate Cancer As Good As a Pathologist.
As Well implies it also 'needs' the Pathologist to complete the set.
Preferably: Researchers suggest AI can diagnose prostate cancer at a professional level.
The methodology (as described in the summary; didn't read TFA) reeks of having testing data mixed in with training data. Did they split the test data out on the patient level, or on the sub sample level? If the latter, the study is garbage.
I can't wait to get a delicate, sensitive robotic probe up my backside. I imagine it'll go something like this .....
Please pull down your pants and bend over. You have twenty seconds to comply
The Russians have won. They have made the world a cesspool of distrust, greed, fear and hate.
It seems that every week we have an 'AI can replace humans' headline only to read the detail and it's only a very very small subset of specifically defined tasks that humans do, and even then only sometimes in very finite cases.
And if anyone knows anything about medicine, half the job is patient comfort and reassurance, something only humans can do.
I wish summaries would include FP and FN rates,
Determination of cancer from images is hardly new - we were showing excellent FP and FN rates for breast cancer in research we were doing ~15 years ago. It's nice to see another useful application, though.
In terms of the use of subdivision of samples, I'd like to see evidence of a clustering of those samples, to ensure there is not hidden information. Given it's peer reviewed, I assume that will have been done.
The long standing advice for men under 70 not to even screen for prostate cancer which , no surprise , seems to be changing again https://edition.cnn.com/2017/04/11/health/prostate-cancer-screening-guidelines-draft/index.html
I did not know that pathologist was a disease, must be one of these new psychiatric conditions that keep popping every now and than. Anyway I'm happy that an AI can diagnose that.
Also, looking forward to see the automated "you got cancer" email, written by an underpaid Indian contractor.
this post contain no useful information, no need to mod it down
Compared to common prostate cancer, mine was a very rare condition that developed from merely lying down.
keep all receipts
Nah, it's just going to create a new class of idiots, robophobes
5. The diagnosis is made. For this until now you need a specialized pathologist with many years of experience and very expensive training to look at the photos and give you his opinion of whether this is cancer he's looking at in the photos or not. Or apparently now you can use a computer program to check out the photos and give you an almost equally accurate opinion.
It's not really like that when you get into the details. My wife is a pathologist. First off anatomic pathologists do not look at photos as a general proposition. They look at slides under a microscope for the most part. There are some pathology imaging systems but they are not in widespread use currently for both cost and technical reasons. It turns out to be technically challenging and expensive in many cases to make an image of every slide with sufficient detail to be useful - it's cheaper and more flexible to just look at the slide directly under a microscope in most cases. Imaging systems will become more common in time but there are a lot of technical issues to work out first.
There are two types of pathology. Clinical pathology and anatomic pathology. Clinical pathology is what is done with stuff like blood draws. It's sent to a lab where the tissue sample is run through some expensive machines which spit out a computer report. You've probably seen some of these. Anatomic pathologists on the other hand deal in tissue samples that result from surgeries or from biopsies. The tissue sample is sent to a lab where it is usually embedded in wax, stained, and then looked at under a microscope to render a diagnosis. This diagnosis is generally based on morphology as well as information gained from molecular and chemical stains. In essence it is pattern matching. In time anatomic pathology will likely come to resemble clinical pathology more and more. However this does not mean the need for the pathologist will go away. It just will mean that their job will involve managing automation and interpreting the results. There are many cases where the results are ambiguous and a human expert will remain necessary to reconcile the problems and interact with the various clinicians to ensure the proper course of treatment occurs.
It is important to understand that the important word in your comment is "opinion". Diagnosis is not binary. Disease criteria are not nearly as well defined as you and I would like them to be. The difference between "mild dysplasia" and "severe dysplasia" is often more of a gestalt thing than a function of rigorous criteria but it can have significant clinical implications (surgery versus no surgery or chemo vs watch and wait). In a lot of diseases there isn't sufficient evidence available to have useful gradations relating to clinical outcomes. It's getting better all the time but there is a lot that is unknown. What the pathologist is doing is essentially making an educated guess based on morphology and other evidence as to what disease processes could be going on. In essence they are being asked to predict how a bit of tissue will grow in the future. They are building a differential diagnosis and explaining which diagnosis they favor and why. Computers can do this and in some cases they can be really helpful in ensuring the differential is complete. But there also will always be those weird and difficult cases which is where it is unlikely the need for humans will go away. Automation will be very helpful to anatomic pathology but it's not going to replace human pathologist any time soon. It will just make them more efficient and (hopefully) reduce costs. What will happen is the computer will spit out a report with some results and a differential but the pathologist will examine the report and interpret it in the
I am very skeptical whether we have any idea who is a good pathologist and who is a bad one.
We do. Pathologists tend to have their work checked a lot (plus it's an accreditation requirement) so other pathologists tend to have a pretty good idea who is good and who isn't. (I know this because my wife is a pathologist) There also is a lot of clinical outcome data out there so it's pretty easy to correlate that to accuracy in diagnosis in pathology. Also if you want to know who is a good clinician the best person to ask is often a pathologist because they get to see the clinicians work. If you want to know who is a good surgeon and who isn't, a pathologist can be the best person to ask.
What you should worry about in pathology though isn't so much whether a given pathologist is good or not but what their incentives are. Like most doctors pathologists are compensated piece rate. Meaning the more cases they look at the more they get paid. There is little to no incentive (aside from avoiding lawsuits) to spend extra time on difficult cases so many of the larger labs crank through ridiculous numbers of cases. This necessarily means that they aren't giving every case their full attention and capability. There should be stronger incentives for outcome based rather than piece rate based compensation. Some labs like my wife's work more carefully but this makes them less profitable.
One important thing to bear in mind is that NO pathologist or doctor of any kind is right 100% of the time. They will make mistakes and there are cases where it will be impossible to get the right diagnosis. It's a bit like playing poker. You make you best guess based on imperfect information and sometimes you're going to get beat because there was something you didn't know.
Accuracy of diagnosis. In case of a pathologist, this isn't about curing people. It's about being correct in one's diagnosis of potentially having specific kinds of cancer.
Careful there. The "correct" diagnosis is only found out by the disease process evolving. What pathologists are trying to find is the diagnosis that meets the standard of care based on the available evidence. Pathologists are essentially being asked to gaze into the future and guess how a disease process will progress. They never have enough information to be right 100% of the time. Its a little like predicting the weather in that regard. So you evaluate pathologist performance by their accuracy in relation to other pathologists looking at the same case with the same information. What you want is the guy with the highest percentage of being right but nobody is going to be right 100% of the time.
You take one sample, split it in two, get a doctor you want to test analyse one sample, get other analysed by trusted expert/panel of experts. Do that across notable amount of samples. You now have success rate for this specific pathologist.
This is done all the time. It's a requirement for accreditation of laboratories not to mention to avoid lawsuits. Difficult cases cases routinely get shown to multiple pathologists even in cases when they aren't trying to track success rates.
General practitioners know everything and do nothing.
Surgeons know nothing and do everything.
Pathologists know everything and do everything but they do it too late.
You are confusing oncologist with pathologist. Pathologists study existing samples for existing conditions.
I'm married to a pathologist and I've worked in and around pathology labs. I assure you that there is no confusion here about what they do. You on the other hand are not actually well informed about what pathologists actually do as evidenced by your comments.
Oncologist is the one that "gazes into the future" to figure out the optimal treatment.
All doctors have to make guesses based on probabilities. The reason we have specialists like pathologists is that medicine is a team activity and you need multiple experts to get the diagnosis in many cases. It's no different than ER doctor calling in a cardiac specialist or a psychiatrist. Pathologists "gaze into the future" to try to predict whether a cell will grow in a manner that will adversely affect the health of the patient. When a pathologist gives a "benign" diagnosis he/she is saying that they do not expect it to be a health risk. If they says something is "malignant" then they are predicting it will be harmful. When they stage and grade lesions they are giving a statistical guess as to the likely course of the disease and the treatments that might be required. It is a guess about the future course of the disease. They have no way to know for certain. It's just an educated guess. Sometimes they are wrong. The reason some diseases like melanoma are so dangerous is that they sometimes mimic benign disease processes.
Now many disease processes cannot be diagnosed by pathology alone or by clinical impression alone. In many cases the diagnosis requires a clinical correlation with pathology so it is de-facto a joint diagnosis and often a joint treatment plan as well. Pathologists routinely prompt treatment modalities through their reports which clinicians ignore at their peril. In their reports they will tell the clinician (in their lingo) "you should cut this lesion out" or "I think this is harmless so no treatment is necessary". While it is the responsibility of the clinician to decide on the treatment, I can assure you that they are heavily guided by the recommendations of the pathologist which often will include recommendations of clinical treatments. If the clinician decides to ignore the pathologists recommendations and things go badly the first thing a lawyer will do is ask why they ignored the pathology report. When a pathologists writes "excision is recommended" that is the pathologist treating the patient and is effectively an order to the clinician to remove tissue. The fact that they aren't holding the scalpel is irrelevant and it is unlikely that such a recommendation will be ignored.
One is the doctor who's job is solely to analyse the sample for existing conditions
You are misinformed as to what a pathologist actually does. Pathologists not only examine tissue for disease processes they also indicate recommendations for treatment as well as a act as reference for potential alternative diagnosis. My wife is asked by the clinicians she works with daily for what she recommends as treatment. (usually a decision regarding whether to surgically remove more tissue) They treat the patient every bit as much as the clinician does and they provide specialist information that would be otherwise unavailable to the clinician.
Other is the doctor that formulates the diagnosis based on, among other things, pathologist's report and then formulate the treatment plan.
Both pathologist and clinicians provide diagnoses and they both matter. Usually the clinician's diagnosis is heavily informed by the pathologist's diagnosis. The clinician does decide on the treatment modality but pathologists routinely tell clinicians what to do. If a pathologist tells a clinician that the diagnosis is melanoma the pathologist is perfectly well aware of what the standard of care for that is and what will
I'm a practicing physician. Here are a two thoughts that come to mind:
1. 99.38% isn't all that great. Most biopsies come back negative. If only 1 out of every 100 biopsies come back positive, you could write a computer program to answer "negative" for each slide and be correct 99% of the time. You have to do some careful statistical analysis to really tell if this sort of AI works- it's not a simple as being right 99.38% of the time.
2. I've never had trouble finding a pathologist to look at a slide. Every hospital has pathologists on staff. It takes longer to turn the biopsy into a set of microscope slides than it does to have the pathologist read them.
I still do not understand how you know that overtreatment is not occurring.
In some cases it certainly is occurring. Sometimes this is appropriate and many others it is not. But it's a more nuanced problem than you might realize.
How do you know treatment is too aggressive?
Several ways but primarily you can compare the pathology to the treatment. If you start seeing treatment disproportionate to expectations for a given diagnosis then you have evidence that over treatment (and by extension over billing) is occurring. This is essentially a statistical evaluation. Also you or the doctor can solicit second (or third) opinions to determine whether the treatment plan is too aggressive.
Now bear in mind that erring slightly on the side of being too aggressive actually can be appropriate. For example a small percentage of appendix removals are expected to actually be unnecessary to ensure that all the ones that should be removed are. That is known and accepted as an appropriate standard of care. The symptoms that would prompt removal can be mimiced by other conditions so doctors cannot be right 100% of the time. Given that a few percent of appendix bursts are fatal, it's better to err on the side of caution. If you have a lesion on your skin that might be melanoma then it's appropriate to remove it even if the chance is small because if you are wrong it can be fatal.
A woman with a tiny lump is rushed under the knife -- lump and nearby lymph nodes are removed. She is immediate given chemo. She does not die of cancer in one year or five years. Maybe that lump was not very dangerous in the first place?
It might be benign but there is no way to know that with absolute certainty in a lot of cases. There are too many variables and unknowns for medicine to be an exact science. Almost all diagnosis are a case of playing the odds. The doctor is making an educated guess based on the probabilities. The doctor sees a lump and the pathologist tells him that 90% of the time it results in a good outcome but 10% of the time it is fatal. No way to know for certain so the proper result is to treat for the worst reasonable outcome. Similarly most of the time if your appendix bursts it will not be fatal. But standard of care is to remove it if certain symptoms appear because a small percentage of the time the outcome is seriously bad. We tend to expect certainty but the reality is that absolute certainty just isn't possible a lot of the time. There is no way to predict with absolute certainty how a dysplastic cell will grow. It's literally impossible to be certain. The best we can do is make an informed guess based on similar conditions we've observed in other patients.
Have there actually been any careful studies where patients with small lumps that are probably cancer but not necessarily dangerous, where some women are treated and some not?
Plenty of them. It's a well studied issue.