Maybe climate change would make certain parts of that area more habitable, though? The sort of climate change that the world experiences won't be uniform.
Individuals with the Delta 32 allele of CCR5 are healthy, suggesting that CCR5 is largely dispensable. However, CCR5 apparently plays a role in mediating resistance to West Nile virus infection in humans, as CCR5-Delta 32 individuals have shown to be disproportionately at higher risk of West Nile virus in studies,[5] indicating that not all of the functions of CCR5 may be compensated by other receptors.
Preventing this receptor from functioning does not appear to disable the immune system, though it may raise the risk of contracting certain specific diseases. This is probably an acceptable tradeoff for the limited period of time a patient with HIV would be required to take such a drug. If you're going to be a jerk, at least make sure you get your information right before you flame.
I put in as many hours of work as I need to get my work done well. That should be the requirement for any job. If that happens to be 3 hours a week, so be it; the system is not challenging me enough to require more of my time. Paying for time spent rewards inefficiency.
Blaming your computer systems for your poor decisions seems to be the new fad of late - it's the second time in a month I've heard this excuse bandied about (was it the state of California that argued it the first time? I forget, but I definitely remember hearing it...)
I agree, but not for the same reasons (although the author's reasoning sounds plausible for big-name journals).
I've found that peer reviewers very seldom give good critiques of the methodology. Rather, most of the comments appear to be on the scope of the paper - comments of the variety "You're doing X. Smith et al. has done Y (which is tangentially and usually very weakly related to X) yet you don't mention this." I suspect that this is because most reviewers don't know enough about the research methods being used to provide a thorough, useful, and accurate critique, but still need to write something, so they take shots at the scope instead, trying to draw on what they do know.
If the methodology is not being effectively critiqued, there is little to no selection for sound research. And if each point you bring up causes reviewers to demand you bring up five more, this introduces selective pressure towards papers that either say far too much or far too little.
I still think a Digg-style system would work well for paper publication. Any half-sound research would be available, but research that people find "useful" would naturally rise to the top.
Animal studies tend to involve some cruel but necessary things. For instance, the biologists on my research team are developing a new targeted therapy for a certain type of tumor. But in order to test it out, they needed to order transgenic mice that are genetically engineered to develop this tumor. They treat the mice well, but it doesn't change the fact that the mice are going to go through a lot of suffering and that their sole purpose in life is to become terminally ill.
I don't like it (and fortunately, as one of the computer scientists on the team, I don't have to deal with it on a regular basis), but I can't see any good alternative. We don't know how safe the therapy is yet.
It is cruel. But it is a necessary evil against the balance of lives it may save.
This is completely unlike any scientific paper I've ever read. It could be more rigorous in its methods, but I wish more papers adopted the same sort of tone. It would make them much more interesting to read.
It is fashionable in general to be anti-intellectual these days.
And it is hard to be an intellectual to begin with. I don't even mean "knowledge worker", I mean outright "creating new mathematics". It takes a lot to think of things that no one has ever discovered before.
Given that mathematics is not an easy field to begin with, and that it also tends to make people terribly unpopular, it's no wonder that we're not doing too well in it. It's very easy to turn most people aside from that sort of career, even if the long-term payoff might be very high.
The question is, does this mean that the remaining mathematicians who defy the pressure are more talented/interested? It would be a small upside.
I didn't read the article very thoroughly, but it sounds like they're comparing distance to the nearest "beautiful" neighbor in their database (according to some distance metric or another), then performing some sort of deformable image registration to try and bring the face closer to that image.
The thing about the nearest neighbor strategy with most distance metrics is that it is reflexive: if you choose an item already in the training set, you are always going to pick that item first. What would happen if they tried to register an image to itself? Presumably nothing, at least within a certain error bound.
If any program succeeds, it is likely to be hailed as the most significant breakthrough in artificial intelligence since the IBM supercomputer Deep Blue beat world chess champion Garry Kasparov in 1997.
I'm still not convinced effective concurrency requires changes to be made in software. I remember AMD working on something that would allow it to be done more or less automatically, even in a single threaded app. I don't know what happened to that or how effective it actually was, but I think the optimal solution is one that is transparent to the programmer, or at least nearly so.
Yes, we're still investigating why the accuracy we obtained was so high. We've ruled out overfitting. I suspect it may be because our images are galactograms (i.e., mammograms with contrast injected into the ducts prior to imaging), which can more easily visualize certain types of abnormalities than unenhanced mammograms. I believe they also come from a single scanner, which isn't usually the case in clinical studies (although I wouldn't expect this to have as much impact in mammography as it does in modalities such as MRI).
Our dataset is smaller than the one used here, but not so small: we had 54 images, 13 of which had tumors. We're correcting for the balance of the classes both by iterated random sampling and by ROC analysis, however, so I'm pretty sure that's not a factor.
Are you referring to the distance metric used in clustering algorithms? Most simply use Euclidean or cosine distance unless they have specific reason to believe a different metric would work better. It isn't really domain dependent.
Or did you mean how "else could you segment a lesion other than by clustering it in the absence of domain knowledge?" There are numerous ways, but edge detection filtering, fuzzy-connectedness segmentation, and watershed segmentation are the first ones that come to mind. They're all general image processing tools.
That's not to say that you couldn't create a clustering algorithm out of those, or that you couldn't combine domain knowledge with these algorithms (the more you know, the better your model can be). But you'd probably get better performance from dedicated clustering algorithms, such as k-means, over using segmentation algorithms in clustering... and perhaps vice versa.
Some people like to call it Computer Aided Risk Estimation (CARE), although some also use this term as a subfield of CAD, but unfortunately, the terminology has become entrenched by this time.
Be careful: A slight improvement in the classifier (or acceptance of another false positive or two) and you may have to make that argument in the other direction. The difference is accuracy is not statistically significant for a binary classification problem of that size.
What this article demonstrates is that current state-of-the-art CAD is nearly as good as a second reader. The performance of the radiologist is pretty much fixed; the algorithm's performance is not.
Clustering algorithms are generally unsupervised. The domain knowledge isn't usually necessary in clustering so much as verifying that the cluster results make sense, since they're much more difficult to quantify than supervised tasks, such as classification, where your data is already labeled.
Of course, you need to do that too before you can present meaningful results and convince people that a system works.
(*As an aside: although you can certainly use a clustering algorithm to segment, the problem you've identified is technically lesion segmentation rather than clustering. There are a lot of non clustering-based approaches to it as well.)
I agree. I actually much prefer working with brains; the organs themselves are more interesting and analyzing the images tends to involve more challenges than 2D mammograms. Volumes vs. static images, spatiotemporal analysis, the option of acquiring functional data to map the lesion to cognitive deficits... I find it a very interesting area. Unfortunately, early diagnosis doesn't always make a difference in certain forms of brain cancer. This needs more research in treatment rather than in diagnosis.
Now we're going into the sociological dynamics of research, which turn out to be really messy, but I'm pretty sure the disproportionate amount of interest in breast cancer is in no small part fueled by the ample funding that gets provided to it vs. other types of cancer. However, as I mentioned in the other post, a lot of the CAD methods tend to be general, and breast cancer is really only a specific application, so this is perhaps not as bad as it sounds (if others apply existing methods elsewhere). Given that other forms of cancer strike more often or have greater mortality rates, and that this one tends to strike only half of the population with any frequency (although it is possible for it to develop in men as well), I think something like pancreatic or colon cancer would be more useful to direct some of the study towards, particularly because the current methods for diagnosis are wholly inadequate in the case of pancreatic cancer and rather invasive in the case of colon cancer.
Prostate cancer may also be a useful cancer to study more due to its high prevalence, but it's also gender-specific and the survival rates are rather high already, so I don't think it would be the first cancer to research on my list.
Most results are presented via ROC curve (for the uninitiated, this is a curve that plots true positive rate against false positive rate based on some threshold for classifying a lesion), so the FPR can theoretically be reduced if you're willing to lose sensitivity as well.
The thing is, the outcomes are not balanced. The risk of missing a cancer is considered far greater than the risk of returning a false positive, so the algorithms are usually created with sensitivity rather than specificity in mind. In my opinion (and since I work on some of these algorithms, my opinion is important:)), this is as it should be, and we should worry about specificity only if we can keep a comparable level of sensitivity.
In any case, the article Yahoo is sourcing from does mention the specificity (which is 1-false positive rate), and it is encouraging: with CAD, the specificity was 96.9%, vs. 97.4% for double reading. Given that sensitivity was also similar (87.2% vs. 87.7%), this article paints CAD in a very favorable light.
Worry not, this is standard practice. Although there is general support that CAD (computer-assisted diagnosis) is effective vs. a second reader, there is still a bit of controversy in the field from time to time, since the results have not been overwhelmingly in favor of CAD yet. There's always at least one talk on the general usefulness of CAD at conferences. Sometimes whole sections get devoted to the topic.
What is a bit more puzzling is why it isn't as prevalent in diagnosis of other types of cancer. Most of the computer-aided detection algorithms draw on general machine learning and image processing techniques rather than specific domain-knowledge of the breast, and thus many of them can be applied, sometimes without any changes, to other organs. There is nothing particularly special about the breast.
My group developed a CAD system for MRI images of the brain, and in the course of performing experiments to put in the paper, I decided to run a few images from a breast CAD project through the classifier. Sure enough, the classifier we had developed for MRIs correctly classified 96% of the mammograms we fed it as well.
Each scientist has a certain balance between open-mindedness and skepticism. I personally favor open-mindedness, even if it comes at the cost of adopting the occasional wrong idea, but I think skepticism is more common. More open-mindedness can improve the uniqueness and number of your ideas, more skepticism their chance of success and the rigor with which they are pursued.
Or, to put it another way, there is an ROC curve for accepting ideas. People who are more skeptical are gaining specificity (less acceptance of what is false) at the cost of sensitivity (less acceptance of what is true too), while less skeptical people trade off in the other direction.
Just being open-minded doesn't mean you automatically accept any idea, though. It still has to make sense. I guess for the GP, this one didn't, although I'm curious why he didn't explicitly state his reasoning.
The probability that the paper trail doesn't match the electronic voting record under an assumption of an unbiased machine is a function of the machine's reliability. If the paper trail is properly kept and the machine truly is reliable enough to use for voting, this probability should be very close to 0. The effect of multiple comparisons on the data should thus be rather small. Still, it may exist, and that's why I backed away somewhat from insisting upon an exact match.
It's late and I'm probably missing something obvious, but I'm a bit unclear on your argument's meaning. Are you arguing that the false positive rate would be too high due to the multiple comparisons you'd have to make to test the sample against the hypothesis? You can tweak the threshold required to identify the result as a mismatch if that's the case. You'll lose sensitivity (so you'll be less likely to detect tampering in the first place) but gain specificity. There's also Bonferroni correction. Another option was suggested by a poster: keep the threshold high and confirm by an examination of a larger sample of the record. Not only could you compare paper trails, but you could also compare the two samples for consistency using something as simple as a t-test, since the mean is an unbiased estimator and you would expect the two samples to have very close means if assuming a lack of systemic bias. (That isn't to say that a latent bias doesn't exist in the voting patterns, but a good random sampling shouldn't have too much of this).
My guess is that this whole discussion is pointless anyway, since we need a verifiable paper trail before any of this would apply and the government has not shown any insistence on one.
Slashdot Mirror
Maybe climate change would make certain parts of that area more habitable, though? The sort of climate change that the world experiences won't be uniform.
From Wikipedia:
Preventing this receptor from functioning does not appear to disable the immune system, though it may raise the risk of contracting certain specific diseases. This is probably an acceptable tradeoff for the limited period of time a patient with HIV would be required to take such a drug. If you're going to be a jerk, at least make sure you get your information right before you flame.
I put in as many hours of work as I need to get my work done well. That should be the requirement for any job. If that happens to be 3 hours a week, so be it; the system is not challenging me enough to require more of my time. Paying for time spent rewards inefficiency.
Blaming your computer systems for your poor decisions seems to be the new fad of late - it's the second time in a month I've heard this excuse bandied about (was it the state of California that argued it the first time? I forget, but I definitely remember hearing it...)
I wonder whether this will help or hurt IT?
I agree, but not for the same reasons (although the author's reasoning sounds plausible for big-name journals).
I've found that peer reviewers very seldom give good critiques of the methodology. Rather, most of the comments appear to be on the scope of the paper - comments of the variety "You're doing X. Smith et al. has done Y (which is tangentially and usually very weakly related to X) yet you don't mention this." I suspect that this is because most reviewers don't know enough about the research methods being used to provide a thorough, useful, and accurate critique, but still need to write something, so they take shots at the scope instead, trying to draw on what they do know.
If the methodology is not being effectively critiqued, there is little to no selection for sound research. And if each point you bring up causes reviewers to demand you bring up five more, this introduces selective pressure towards papers that either say far too much or far too little.
I still think a Digg-style system would work well for paper publication. Any half-sound research would be available, but research that people find "useful" would naturally rise to the top.
Animal studies tend to involve some cruel but necessary things. For instance, the biologists on my research team are developing a new targeted therapy for a certain type of tumor. But in order to test it out, they needed to order transgenic mice that are genetically engineered to develop this tumor. They treat the mice well, but it doesn't change the fact that the mice are going to go through a lot of suffering and that their sole purpose in life is to become terminally ill.
I don't like it (and fortunately, as one of the computer scientists on the team, I don't have to deal with it on a regular basis), but I can't see any good alternative. We don't know how safe the therapy is yet.
It is cruel. But it is a necessary evil against the balance of lives it may save.
Found it.
The author and year alone are not enough, but couple it with the subject of the research and finding it is fairly trivial.
He did publish a couple of other studies on this during 2006, however, so this may still not be the study the GP was citing.
This is completely unlike any scientific paper I've ever read. It could be more rigorous in its methods, but I wish more papers adopted the same sort of tone. It would make them much more interesting to read.
It is fashionable in general to be anti-intellectual these days.
And it is hard to be an intellectual to begin with. I don't even mean "knowledge worker", I mean outright "creating new mathematics". It takes a lot to think of things that no one has ever discovered before.
Given that mathematics is not an easy field to begin with, and that it also tends to make people terribly unpopular, it's no wonder that we're not doing too well in it. It's very easy to turn most people aside from that sort of career, even if the long-term payoff might be very high.
The question is, does this mean that the remaining mathematicians who defy the pressure are more talented/interested? It would be a small upside.
I didn't read the article very thoroughly, but it sounds like they're comparing distance to the nearest "beautiful" neighbor in their database (according to some distance metric or another), then performing some sort of deformable image registration to try and bring the face closer to that image.
The thing about the nearest neighbor strategy with most distance metrics is that it is reflexive: if you choose an item already in the training set, you are always going to pick that item first. What would happen if they tried to register an image to itself? Presumably nothing, at least within a certain error bound.
And just as intelligent, to boot.
I'm still not convinced effective concurrency requires changes to be made in software. I remember AMD working on something that would allow it to be done more or less automatically, even in a single threaded app. I don't know what happened to that or how effective it actually was, but I think the optimal solution is one that is transparent to the programmer, or at least nearly so.
Yeah, it's technically an oblate spheroid.
If we can only find them 2 hours before they're scheduled to hit, that's about all we can do.
Yes, we're still investigating why the accuracy we obtained was so high. We've ruled out overfitting. I suspect it may be because our images are galactograms (i.e., mammograms with contrast injected into the ducts prior to imaging), which can more easily visualize certain types of abnormalities than unenhanced mammograms. I believe they also come from a single scanner, which isn't usually the case in clinical studies (although I wouldn't expect this to have as much impact in mammography as it does in modalities such as MRI).
Our dataset is smaller than the one used here, but not so small: we had 54 images, 13 of which had tumors. We're correcting for the balance of the classes both by iterated random sampling and by ROC analysis, however, so I'm pretty sure that's not a factor.
Are you referring to the distance metric used in clustering algorithms? Most simply use Euclidean or cosine distance unless they have specific reason to believe a different metric would work better. It isn't really domain dependent.
Or did you mean how "else could you segment a lesion other than by clustering it in the absence of domain knowledge?" There are numerous ways, but edge detection filtering, fuzzy-connectedness segmentation, and watershed segmentation are the first ones that come to mind. They're all general image processing tools.
That's not to say that you couldn't create a clustering algorithm out of those, or that you couldn't combine domain knowledge with these algorithms (the more you know, the better your model can be). But you'd probably get better performance from dedicated clustering algorithms, such as k-means, over using segmentation algorithms in clustering... and perhaps vice versa.
*"Volumes vs. 2D images", that is.
Some people like to call it Computer Aided Risk Estimation (CARE), although some also use this term as a subfield of CAD, but unfortunately, the terminology has become entrenched by this time.
Be careful: A slight improvement in the classifier (or acceptance of another false positive or two) and you may have to make that argument in the other direction. The difference is accuracy is not statistically significant for a binary classification problem of that size.
What this article demonstrates is that current state-of-the-art CAD is nearly as good as a second reader. The performance of the radiologist is pretty much fixed; the algorithm's performance is not.
Clustering algorithms are generally unsupervised. The domain knowledge isn't usually necessary in clustering so much as verifying that the cluster results make sense, since they're much more difficult to quantify than supervised tasks, such as classification, where your data is already labeled.
Of course, you need to do that too before you can present meaningful results and convince people that a system works.
(*As an aside: although you can certainly use a clustering algorithm to segment, the problem you've identified is technically lesion segmentation rather than clustering. There are a lot of non clustering-based approaches to it as well.)
I fell right into that one, didn't I? :)
I agree. I actually much prefer working with brains; the organs themselves are more interesting and analyzing the images tends to involve more challenges than 2D mammograms. Volumes vs. static images, spatiotemporal analysis, the option of acquiring functional data to map the lesion to cognitive deficits... I find it a very interesting area. Unfortunately, early diagnosis doesn't always make a difference in certain forms of brain cancer. This needs more research in treatment rather than in diagnosis.
Now we're going into the sociological dynamics of research, which turn out to be really messy, but I'm pretty sure the disproportionate amount of interest in breast cancer is in no small part fueled by the ample funding that gets provided to it vs. other types of cancer. However, as I mentioned in the other post, a lot of the CAD methods tend to be general, and breast cancer is really only a specific application, so this is perhaps not as bad as it sounds (if others apply existing methods elsewhere). Given that other forms of cancer strike more often or have greater mortality rates, and that this one tends to strike only half of the population with any frequency (although it is possible for it to develop in men as well), I think something like pancreatic or colon cancer would be more useful to direct some of the study towards, particularly because the current methods for diagnosis are wholly inadequate in the case of pancreatic cancer and rather invasive in the case of colon cancer.
Prostate cancer may also be a useful cancer to study more due to its high prevalence, but it's also gender-specific and the survival rates are rather high already, so I don't think it would be the first cancer to research on my list.
Most results are presented via ROC curve (for the uninitiated, this is a curve that plots true positive rate against false positive rate based on some threshold for classifying a lesion), so the FPR can theoretically be reduced if you're willing to lose sensitivity as well.
The thing is, the outcomes are not balanced. The risk of missing a cancer is considered far greater than the risk of returning a false positive, so the algorithms are usually created with sensitivity rather than specificity in mind. In my opinion (and since I work on some of these algorithms, my opinion is important :)), this is as it should be, and we should worry about specificity only if we can keep a comparable level of sensitivity.
In any case, the article Yahoo is sourcing from does mention the specificity (which is 1-false positive rate), and it is encouraging: with CAD, the specificity was 96.9%, vs. 97.4% for double reading. Given that sensitivity was also similar (87.2% vs. 87.7%), this article paints CAD in a very favorable light.
Worry not, this is standard practice. Although there is general support that CAD (computer-assisted diagnosis) is effective vs. a second reader, there is still a bit of controversy in the field from time to time, since the results have not been overwhelmingly in favor of CAD yet. There's always at least one talk on the general usefulness of CAD at conferences. Sometimes whole sections get devoted to the topic.
What is a bit more puzzling is why it isn't as prevalent in diagnosis of other types of cancer. Most of the computer-aided detection algorithms draw on general machine learning and image processing techniques rather than specific domain-knowledge of the breast, and thus many of them can be applied, sometimes without any changes, to other organs. There is nothing particularly special about the breast.
My group developed a CAD system for MRI images of the brain, and in the course of performing experiments to put in the paper, I decided to run a few images from a breast CAD project through the classifier. Sure enough, the classifier we had developed for MRIs correctly classified 96% of the mammograms we fed it as well.
Each scientist has a certain balance between open-mindedness and skepticism. I personally favor open-mindedness, even if it comes at the cost of adopting the occasional wrong idea, but I think skepticism is more common. More open-mindedness can improve the uniqueness and number of your ideas, more skepticism their chance of success and the rigor with which they are pursued.
Or, to put it another way, there is an ROC curve for accepting ideas. People who are more skeptical are gaining specificity (less acceptance of what is false) at the cost of sensitivity (less acceptance of what is true too), while less skeptical people trade off in the other direction.
Just being open-minded doesn't mean you automatically accept any idea, though. It still has to make sense. I guess for the GP, this one didn't, although I'm curious why he didn't explicitly state his reasoning.
The probability that the paper trail doesn't match the electronic voting record under an assumption of an unbiased machine is a function of the machine's reliability. If the paper trail is properly kept and the machine truly is reliable enough to use for voting, this probability should be very close to 0. The effect of multiple comparisons on the data should thus be rather small. Still, it may exist, and that's why I backed away somewhat from insisting upon an exact match.
It's late and I'm probably missing something obvious, but I'm a bit unclear on your argument's meaning. Are you arguing that the false positive rate would be too high due to the multiple comparisons you'd have to make to test the sample against the hypothesis? You can tweak the threshold required to identify the result as a mismatch if that's the case. You'll lose sensitivity (so you'll be less likely to detect tampering in the first place) but gain specificity. There's also Bonferroni correction. Another option was suggested by a poster: keep the threshold high and confirm by an examination of a larger sample of the record. Not only could you compare paper trails, but you could also compare the two samples for consistency using something as simple as a t-test, since the mean is an unbiased estimator and you would expect the two samples to have very close means if assuming a lack of systemic bias. (That isn't to say that a latent bias doesn't exist in the voting patterns, but a good random sampling shouldn't have too much of this).
My guess is that this whole discussion is pointless anyway, since we need a verifiable paper trail before any of this would apply and the government has not shown any insistence on one.