My main concentration so far has been focusing on unsupervised training for identification of the 2 frames (for any arbitrary slice) pertaining to end-systolic and end-diastolic stages of the heart beat cadence.
\n\nWe know that these two stages are defined as the global minimum and maximum of the entire left ventrical volume, thus it seems logical to assume that these global maximums and minimums (for volume) ALSO occur at the global maximum and minimum (for area) of the individual cross-sections for a given slice.
\n\nHowever, one can imagine a situation where this might not necessarily be the case - eg. at maximum LV volume, one LV cross section area is slightly smaller than it is, say, slightly after maximum LV volume.
\n\nI have no reason to believe that this is the case, and one would certainly think that maximum volume occurs at the maximum cross-section areas for each slice.
\n\nWhat are peoples thoughts? Any anatomy experts that could shed some light on wether this is likely to be true or not?
\n\nIs anyone else tackling this problem in a similar fashion (first identifying 2 frames from each slice, in which to build a more complex model)? My gut reaction is that it doesn't make sense to use more than 2 frames per slice.
\n\nI have been having success training restricted boltzmann machines with 1 hidden layer and 1 node (essentially what amounts to a logistic regression) on slices. This works out well because we can do it totally unsupervised - we have no labeling pertaining to the cadence. Essentially, we want to train a model to capture a generative distribution for a given slice MRI. The hope is that the single binary output of the model pertains to end-systolic and end-diastolic stages in the slice.
\n\nAs an example, I trained an RBM on slice 9 of study 525 (validation data), by first finding the 128x128 window within the whole image that has the most variance across the frames.
\n\nAfter training on the grey-scale image, we end up with the following weight:\nweight
\n\nWhich when "off" will decode to:\noff
\n\nAnd when "on" will decode to:\non
\n\nThis confirms that we have successfully trained an auto-encoding binary feature pertaining to heart beat cadence, and completely unsupervised! This is the power of generative distributions.
\n\nNow keep in mind, all we have done here is create a model which is good at assigning a 1 or a 0 to each frame in a way that attempts to capture as much information as possible about the differences between frames. What is ultimately more useful is the dot product of our image vector and the trained weight (essentially, what we would plug in to the sigmoid function). Calculating that dot-product for each frame gives us this beautiful curve:\ncurve
\n\nSo it seems our model is indicating that frame 13 and 29 are our likeliest candidates for the end-systolic and end-diastolic stages in the slice. Checking out the whole slice confirms that this is correct:\nslice
\n\nWhadaya think? :)
", "messages": [ { "id": "102181", "postDate": "12/20/2015 08:06:32", "content": "My main concentration so far has been focusing on unsupervised training for identification of the 2 frames (for any arbitrary slice) pertaining to end-systolic and end-diastolic stages of the heart beat cadence.
\n\nWe know that these two stages are defined as the global minimum and maximum of the entire left ventrical volume, thus it seems logical to assume that these global maximums and minimums (for volume) ALSO occur at the global maximum and minimum (for area) of the individual cross-sections for a given slice.
\n\nHowever, one can imagine a situation where this might not necessarily be the case - eg. at maximum LV volume, one LV cross section area is slightly smaller than it is, say, slightly after maximum LV volume.
\n\nI have no reason to believe that this is the case, and one would certainly think that maximum volume occurs at the maximum cross-section areas for each slice.
\n\nWhat are peoples thoughts? Any anatomy experts that could shed some light on wether this is likely to be true or not?
\n\nIs anyone else tackling this problem in a similar fashion (first identifying 2 frames from each slice, in which to build a more complex model)? My gut reaction is that it doesn't make sense to use more than 2 frames per slice.
\n\nI have been having success training restricted boltzmann machines with 1 hidden layer and 1 node (essentially what amounts to a logistic regression) on slices. This works out well because we can do it totally unsupervised - we have no labeling pertaining to the cadence. Essentially, we want to train a model to capture a generative distribution for a given slice MRI. The hope is that the single binary output of the model pertains to end-systolic and end-diastolic stages in the slice.
\n\nAs an example, I trained an RBM on slice 9 of study 525 (validation data), by first finding the 128x128 window within the whole image that has the most variance across the frames.
\n\nAfter training on the grey-scale image, we end up with the following weight:\nweight
\n\nWhich when "off" will decode to:\noff
\n\nAnd when "on" will decode to:\non
\n\nThis confirms that we have successfully trained an auto-encoding binary feature pertaining to heart beat cadence, and completely unsupervised! This is the power of generative distributions.
\n\nNow keep in mind, all we have done here is create a model which is good at assigning a 1 or a 0 to each frame in a way that attempts to capture as much information as possible about the differences between frames. What is ultimately more useful is the dot product of our image vector and the trained weight (essentially, what we would plug in to the sigmoid function). Calculating that dot-product for each frame gives us this beautiful curve:\ncurve
\n\nSo it seems our model is indicating that frame 13 and 29 are our likeliest candidates for the end-systolic and end-diastolic stages in the slice. Checking out the whole slice confirms that this is correct:\nslice
\n\nWhadaya think? :)
", "rawMarkdown": "My main concentration so far has been focusing on unsupervised training for identification of the 2 frames (for any arbitrary slice) pertaining to end-systolic and end-diastolic stages of the heart beat cadence.\r\n\r\nWe know that these two stages are defined as the global minimum and maximum of the entire left ventrical volume, thus it seems logical to assume that these global maximums and minimums (for volume) ALSO occur at the global maximum and minimum (for area) of the individual cross-sections for a given slice.\r\n\r\nHowever, one can *imagine* a situation where this might not necessarily be the case - eg. at maximum LV volume, one LV cross section area is slightly smaller than it is, say, slightly after maximum LV volume.\r\n\r\nI have no reason to believe that this is the case, and one would certainly think that maximum volume occurs at the maximum cross-section areas for each slice.\r\n\r\nWhat are peoples thoughts? Any anatomy experts that could shed some light on wether this is likely to be true or not?\r\n\r\nIs anyone else tackling this problem in a similar fashion (first identifying 2 frames from each slice, in which to build a more complex model)? My gut reaction is that it doesn't make sense to use more than 2 frames per slice.\r\n\r\nI have been having success training restricted boltzmann machines with 1 hidden layer and 1 node (essentially what amounts to a logistic regression) on slices. This works out well because we can do it totally unsupervised - we have no labeling pertaining to the cadence. Essentially, we want to train a model to capture a generative distribution for a given slice MRI. The hope is that the single binary output of the model pertains to end-systolic and end-diastolic stages in the slice. \r\n\r\nAs an example, I trained an RBM on slice 9 of study 525 (validation data), by first finding the 128x128 window within the whole image that has the most variance across the frames.\r\n\r\nAfter training on the grey-scale image, we end up with the following weight:\r\n[weight][2]\r\n\r\nWhich when \"off\" will decode to:\r\n[off][3]\r\n\r\nAnd when \"on\" will decode to:\r\n[on][4]\r\n\r\nThis confirms that we have successfully trained an auto-encoding binary feature pertaining to heart beat cadence, and *completely* unsupervised! This is the power of generative distributions.\r\n\r\nNow keep in mind, all we have done here is create a model which is good at assigning a 1 or a 0 to each frame in a way that attempts to capture as much information as possible about the differences between frames. What is ultimately more useful is the dot product of our image vector and the trained weight (essentially, what we would plug in to the sigmoid function). Calculating that dot-product for each frame gives us this beautiful curve:\r\n[curve][5]\r\n\r\nSo it seems our model is indicating that frame 13 and 29 are our likeliest candidates for the end-systolic and end-diastolic stages in the slice. Checking out the whole slice confirms that this is correct:\r\n[slice][6]\r\n\r\nWhadaya think? :)\r\n\r\n [2]: http://imgur.com/zso6tx7\r\n [3]: http://imgur.com/d1gKeOt\r\n [4]: http://imgur.com/hCx8mP8\r\n [5]: http://imgur.com/KlZTXEt\r\n [6]: http://imgur.com/fxbEsTE", "votes": null }, { "id": "102210", "postDate": "12/20/2015 13:14:57", "content": "I may be able to add a brief comment regarding slice synchrony.
\n\nIf I understand your question correctly, you are asking if the timing of systole (smallest volume) and the timing of the smallest cross-sectional area for each slice is always the same?
\n\nThe short answer is no. In a healthy heart, it would be mostly the case. There can be small differences in timing as the electrical depolarization of the muscle cells spreads across the heart, but in a healthy heart, contraction is mostly synchronous across the slices.
\n\nI believe the correct term for the pathological condition where timing is off is "intraventricular dyssynchrony", where there is delay or heterogeneity in the timing of contraction in different myocardial segments within the left ventricle. If you google that, you should find more information.
\n\nIntraventricular dyssynchrony would obviously affect cardiac function. Poor synchronization leads (in general) to poor pumping function (i.e., a low ejection fraction).
\n\nHope this helps.
", "rawMarkdown": "I may be able to add a brief comment regarding slice synchrony. \r\n\r\nIf I understand your question correctly, you are asking if the timing of systole (smallest volume) and the timing of the smallest cross-sectional area for each slice is always the same?\r\n\r\nThe short answer is no. In a healthy heart, it would be mostly the case. There can be small differences in timing as the electrical depolarization of the muscle cells spreads across the heart, but in a healthy heart, contraction is mostly synchronous across the slices. \r\n\r\nI believe the correct term for the pathological condition where timing is off is \"intraventricular dyssynchrony\", where there is delay or heterogeneity in the timing of contraction in different myocardial segments within the left ventricle. If you google that, you should find more information. \r\n\r\nIntraventricular dyssynchrony would obviously affect cardiac function. Poor synchronization leads (in general) to poor pumping function (i.e., a low ejection fraction).\r\n\r\nHope this helps.", "votes": null }, { "id": "102234", "postDate": "12/20/2015 19:37:24", "content": "[quote=Michael Hansen;102210]
\n\nI may be able to add a brief comment regarding slice synchrony.
\n\nIf I understand your question correctly, you are asking if the timing of systole (smallest volume) and the timing of the smallest cross-sectional area for each slice is always the same?
\n\nThe short answer is no. In a healthy heart, it would be mostly the case. There can be small differences in timing as the electrical depolarization of the muscle cells spreads across the heart, but in a healthy heart, contraction is mostly synchronous across the slices.
\n\nI believe the correct term for the pathological condition where timing is off is "intraventricular dyssynchrony", where there is delay or heterogeneity in the timing of contraction in different myocardial segments within the left ventricle. If you google that, you should find more information.
\n\nIntraventricular dyssynchrony would obviously affect cardiac function. Poor synchronization leads (in general) to poor pumping function (i.e., a low ejection fraction).
\n\nHope this helps.
\n\n[/quote]\nThis is exactly what I was looking for.
\n\nThanks so much Michael!
", "rawMarkdown": "[quote=Michael Hansen;102210]\r\n\r\nI may be able to add a brief comment regarding slice synchrony. \r\n\r\nIf I understand your question correctly, you are asking if the timing of systole (smallest volume) and the timing of the smallest cross-sectional area for each slice is always the same?\r\n\r\nThe short answer is no. In a healthy heart, it would be mostly the case. There can be small differences in timing as the electrical depolarization of the muscle cells spreads across the heart, but in a healthy heart, contraction is mostly synchronous across the slices. \r\n\r\nI believe the correct term for the pathological condition where timing is off is \"intraventricular dyssynchrony\", where there is delay or heterogeneity in the timing of contraction in different myocardial segments within the left ventricle. If you google that, you should find more information. \r\n\r\nIntraventricular dyssynchrony would obviously affect cardiac function. Poor synchronization leads (in general) to poor pumping function (i.e., a low ejection fraction).\r\n\r\nHope this helps.\r\n\r\n[/quote]\r\nThis is *exactly* what I was looking for.\r\n\r\nThanks so much Michael!", "votes": null }, { "id": "102314", "postDate": "12/21/2015 14:05:44", "content": "Here is an additional reply from, Andrew Arai:
\n\n\n\n\nI’d suggest a slightly different answer on the question about finding end systole and end diastole.
\n \nIn general, the cine MRI is acquired with a consistent number of\n images across the cardiac cycle. The acquisition is triggered so as to\n coordinate the acquisition with the cardiac cycle. Thus, in general,\n end diastole and end systole should be at a similar time frame across\n the cardiac cycle for all slices.
\n \nThat assumption is not perfect since there may be heart rate\n variability due to biological variations from breath hold to breath\n hold and may vary slightly as a function of time. So for measurements\n made at the NIH, we pick a column of images at the same end diastolic\n and end systolic time frame even though we may see that for a given\n slice that one frame earlier or later might be better. Thus, the\n reference standard was measured based on a consistent end diastolic\n and end systolic time frame for the whole short axis stack of images.
\n \nFrom a functional perspective, there are disease states where end\n diastole and end systole do not correspond to the minimum and maximum\n cross sectional area for some slices. For example, if a patient has a\n heart attack, most of the slices should have a smaller end systolic\n and larger end diastolic volume. However, slices that are dominated by\n dysfunctional myocardium associated with the heart attack may be\n dyskinetic. Dyskinesis means a segment bulges while other parts of the\n heart are contracting. Therefore, those slices may have a larger cross\n sectional area at end systole than end diastole.
\n \nTo bring this to a close, I’d personally recommend thinking about the\n finding end systole based on the smallest LV volume and end diastole\n as the largest LV volume as opposed to thinking only slice by slice.
\n
(Drs. Michael Hansen and Andrew Arai are colleagues at NIH NHLBI, and are the PIs for the Data Science Bowl)
", "rawMarkdown": "Here is an additional reply from, Andrew Arai:\r\n\r\n\r\n> I’d suggest a slightly different answer on the question about finding end systole and end diastole.\r\n> \r\n> In general, the cine MRI is acquired with a consistent number of\r\n> images across the cardiac cycle. The acquisition is triggered so as to\r\n> coordinate the acquisition with the cardiac cycle. Thus, in general,\r\n> end diastole and end systole should be at a similar time frame across\r\n> the cardiac cycle for all slices. \r\n> \r\n> That assumption is not perfect since there may be heart rate\r\n> variability due to biological variations from breath hold to breath\r\n> hold and may vary slightly as a function of time. So for measurements\r\n> made at the NIH, we pick a column of images at the same end diastolic\r\n> and end systolic time frame even though we may see that for a given\r\n> slice that one frame earlier or later might be better. Thus, the\r\n> reference standard was measured based on a consistent end diastolic\r\n> and end systolic time frame for the whole short axis stack of images.\r\n> \r\n> From a functional perspective, there are disease states where end\r\n> diastole and end systole do not correspond to the minimum and maximum\r\n> cross sectional area for some slices. For example, if a patient has a\r\n> heart attack, most of the slices should have a smaller end systolic\r\n> and larger end diastolic volume. However, slices that are dominated by\r\n> dysfunctional myocardium associated with the heart attack may be\r\n> dyskinetic. Dyskinesis means a segment bulges while other parts of the\r\n> heart are contracting. Therefore, those slices may have a larger cross\r\n> sectional area at end systole than end diastole.\r\n> \r\n> To bring this to a close, I’d personally recommend thinking about the\r\n> finding end systole based on the smallest LV volume and end diastole\r\n> as the largest LV volume as opposed to thinking only slice by slice.\r\n\r\n(Drs. Michael Hansen and Andrew Arai are colleagues at NIH NHLBI, and are the PIs for the Data Science Bowl)", "votes": null }, { "id": "102368", "postDate": "12/21/2015 22:44:08", "content": "Thank you so much to both PIs for the incredibly thoughtful responses alluding to both the practice that the NIH uses, as well as situations whereby frames may not be synchronized (though they generally are), and situations in which maximum and minimum cross-sectional area is not necessarily correlated with end systole and end diastole (though they generally are). This is incredibly helpful information to have for machine-learning practitioners who are not experts in cardiology!
\n\nRegarding Andrew Arai's last statement:
\n\n[quote=Shannon;102314]
\n\n\n\n\nTo bring this to a close, I’d personally recommend thinking about the\n finding end systole based on the smallest LV volume and end diastole\n as the largest LV volume as opposed to thinking only slice by slice.\n [/quote]
\n
My concern is that the logic here seems to be reversed. Andrew is suggesting that we calculate LV volumes, and use this to determine end systole and end diastole. This challenge is essentially about doing the reverse - (ostensibly) determining which frames pertain to end systole and end diastole for a given slice, knowing that these frames necessarily are the best predictors, and using these frames to calculate LV volumes. I'm unsure of what is being suggested here, any clarification would help greatly.
\n\nThis brings to light another question - how have our ground truth volumes been calculated? Have they been calculated by cardiologists based on MRIs, or have they been calculated using some form of instrumentation which yields true values?
\n\nAgain thank you so much to both PIs for being incredibly involved in this discussion, it is absolutely invaluable!
\n\nCheers
", "rawMarkdown": "Thank you so much to both PIs for the incredibly thoughtful responses alluding to both the practice that the NIH uses, as well as situations whereby frames may not be synchronized (though they generally are), and situations in which maximum and minimum cross-sectional area is not necessarily correlated with end systole and end diastole (though they generally are). This is incredibly helpful information to have for machine-learning practitioners who are not experts in cardiology!\r\n\r\nRegarding Andrew Arai's last statement:\r\n\r\n[quote=Shannon;102314]\r\n> To bring this to a close, I’d personally recommend thinking about the\r\n> finding end systole based on the smallest LV volume and end diastole\r\n> as the largest LV volume as opposed to thinking only slice by slice.\r\n[/quote]\r\n\r\nMy concern is that the logic here seems to be reversed. Andrew is suggesting that we calculate LV volumes, and use this to determine end systole and end diastole. This challenge is essentially about doing the reverse - (ostensibly) determining which frames pertain to end systole and end diastole for a given slice, knowing that these frames necessarily are the best predictors, and using these frames to calculate LV volumes. I'm unsure of what is being suggested here, any clarification would help greatly.\r\n\r\nThis brings to light another question - how have our ground truth volumes been calculated? Have they been calculated by cardiologists based on MRIs, or have they been calculated using some form of instrumentation which yields true values?\r\n\r\nAgain thank you so much to both PIs for being incredibly involved in this discussion, it is absolutely invaluable!\r\n\r\nCheers", "votes": null }, { "id": "102377", "postDate": "12/21/2015 23:34:15", "content": "Let me see if I can help.
\n\nThere is really nothing reversed about the logic. One way to approach is, as you say, estimate the volume for every time point of the cardiac cycle and report the smallest volume as LVESV and the largest volume as LVEDV. In current clinical practice this is too time consuming because it requires manually tracing the ventricles at lots of time points, so the clinician will eyeball which phases have the largest and smallest volume and then segment only those. It would be much better if the volume was calculated for every phase. It is up to you whether your algorithm estimates which phases are systole and diastole and then calculate the volume in those phases or you calculate the volume for all phases and then report min and max. The latter approach is certainly much simpler conceptually and I would prefer an algorithm that does that since the time course has diagnostic value too, but that is not the topic of the competition.
\n\nGround truth for this competition has been calculated based on hand drawn (by a cardiologist) contours. If we had some instrumentation that yields "true" values, we would just use that.
\n\nHope this helps.
", "rawMarkdown": "Let me see if I can help.\r\n\r\nThere is really nothing reversed about the logic. One way to approach is, as you say, estimate the volume for every time point of the cardiac cycle and report the smallest volume as LVESV and the largest volume as LVEDV. In current clinical practice this is too time consuming because it requires manually tracing the ventricles at lots of time points, so the clinician will eyeball which phases have the largest and smallest volume and then segment only those. It would be much better if the volume was calculated for every phase. It is up to you whether your algorithm estimates which phases are systole and diastole and then calculate the volume in those phases or you calculate the volume for all phases and then report min and max. The latter approach is certainly much simpler conceptually and I would prefer an algorithm that does that since the time course has diagnostic value too, but that is not the topic of the competition. \r\n\r\nGround truth for this competition has been calculated based on hand drawn (by a cardiologist) contours. If we had some instrumentation that yields \"true\" values, we would just use that. \r\n\r\nHope this helps.", "votes": null }, { "id": "103948", "postDate": "01/08/2016 04:32:36", "content": "Hi Martin,
\n\nDo you need an RBM network to find smallest and largest cross-sectional areas in slices? Can't you diff successive frames, and then see which ones are non-zeros, and then arrive at cross-sectional areas for slices? Just thinking whether this diffing approach is a viable one or not.
\n\nI am thinking diffing approach is easier and gets ai and bi of simpson's formula also easily.
", "rawMarkdown": "Hi Martin, \r\n\r\nDo you need an RBM network to find smallest and largest cross-sectional areas in slices? Can't you diff successive frames, and then see which ones are non-zeros, and then arrive at cross-sectional areas for slices? Just thinking whether this diffing approach is a viable one or not. \r\n\r\nI am thinking diffing approach is easier and gets ai and bi of simpson's formula also easily.", "votes": null } ], "comments": [ { "id": 102210, "author_name": "michaelhansen", "author_url": "", "post_date": "12/20/2015 13:14:57", "content": "I may be able to add a brief comment regarding slice synchrony.
\n\nIf I understand your question correctly, you are asking if the timing of systole (smallest volume) and the timing of the smallest cross-sectional area for each slice is always the same?
\n\nThe short answer is no. In a healthy heart, it would be mostly the case. There can be small differences in timing as the electrical depolarization of the muscle cells spreads across the heart, but in a healthy heart, contraction is mostly synchronous across the slices.
\n\nI believe the correct term for the pathological condition where timing is off is "intraventricular dyssynchrony", where there is delay or heterogeneity in the timing of contraction in different myocardial segments within the left ventricle. If you google that, you should find more information.
\n\nIntraventricular dyssynchrony would obviously affect cardiac function. Poor synchronization leads (in general) to poor pumping function (i.e., a low ejection fraction).
\n\nHope this helps.
", "votes": null, "replies": [] }, { "id": 102234, "author_name": "mnemaric", "author_url": "", "post_date": "12/20/2015 19:37:24", "content": "[quote=Michael Hansen;102210]
\n\nI may be able to add a brief comment regarding slice synchrony.
\n\nIf I understand your question correctly, you are asking if the timing of systole (smallest volume) and the timing of the smallest cross-sectional area for each slice is always the same?
\n\nThe short answer is no. In a healthy heart, it would be mostly the case. There can be small differences in timing as the electrical depolarization of the muscle cells spreads across the heart, but in a healthy heart, contraction is mostly synchronous across the slices.
\n\nI believe the correct term for the pathological condition where timing is off is "intraventricular dyssynchrony", where there is delay or heterogeneity in the timing of contraction in different myocardial segments within the left ventricle. If you google that, you should find more information.
\n\nIntraventricular dyssynchrony would obviously affect cardiac function. Poor synchronization leads (in general) to poor pumping function (i.e., a low ejection fraction).
\n\nHope this helps.
\n\n[/quote]\nThis is exactly what I was looking for.
\n\nThanks so much Michael!
", "votes": null, "replies": [] }, { "id": 102314, "author_name": "shannonlantzy", "author_url": "", "post_date": "12/21/2015 14:05:44", "content": "Here is an additional reply from, Andrew Arai:
\n\n\n\n\nI’d suggest a slightly different answer on the question about finding end systole and end diastole.
\n \nIn general, the cine MRI is acquired with a consistent number of\n images across the cardiac cycle. The acquisition is triggered so as to\n coordinate the acquisition with the cardiac cycle. Thus, in general,\n end diastole and end systole should be at a similar time frame across\n the cardiac cycle for all slices.
\n \nThat assumption is not perfect since there may be heart rate\n variability due to biological variations from breath hold to breath\n hold and may vary slightly as a function of time. So for measurements\n made at the NIH, we pick a column of images at the same end diastolic\n and end systolic time frame even though we may see that for a given\n slice that one frame earlier or later might be better. Thus, the\n reference standard was measured based on a consistent end diastolic\n and end systolic time frame for the whole short axis stack of images.
\n \nFrom a functional perspective, there are disease states where end\n diastole and end systole do not correspond to the minimum and maximum\n cross sectional area for some slices. For example, if a patient has a\n heart attack, most of the slices should have a smaller end systolic\n and larger end diastolic volume. However, slices that are dominated by\n dysfunctional myocardium associated with the heart attack may be\n dyskinetic. Dyskinesis means a segment bulges while other parts of the\n heart are contracting. Therefore, those slices may have a larger cross\n sectional area at end systole than end diastole.
\n \nTo bring this to a close, I’d personally recommend thinking about the\n finding end systole based on the smallest LV volume and end diastole\n as the largest LV volume as opposed to thinking only slice by slice.
\n
(Drs. Michael Hansen and Andrew Arai are colleagues at NIH NHLBI, and are the PIs for the Data Science Bowl)
", "votes": null, "replies": [] }, { "id": 102368, "author_name": "mnemaric", "author_url": "", "post_date": "12/21/2015 22:44:08", "content": "Thank you so much to both PIs for the incredibly thoughtful responses alluding to both the practice that the NIH uses, as well as situations whereby frames may not be synchronized (though they generally are), and situations in which maximum and minimum cross-sectional area is not necessarily correlated with end systole and end diastole (though they generally are). This is incredibly helpful information to have for machine-learning practitioners who are not experts in cardiology!
\n\nRegarding Andrew Arai's last statement:
\n\n[quote=Shannon;102314]
\n\n\n\n\nTo bring this to a close, I’d personally recommend thinking about the\n finding end systole based on the smallest LV volume and end diastole\n as the largest LV volume as opposed to thinking only slice by slice.\n [/quote]
\n
My concern is that the logic here seems to be reversed. Andrew is suggesting that we calculate LV volumes, and use this to determine end systole and end diastole. This challenge is essentially about doing the reverse - (ostensibly) determining which frames pertain to end systole and end diastole for a given slice, knowing that these frames necessarily are the best predictors, and using these frames to calculate LV volumes. I'm unsure of what is being suggested here, any clarification would help greatly.
\n\nThis brings to light another question - how have our ground truth volumes been calculated? Have they been calculated by cardiologists based on MRIs, or have they been calculated using some form of instrumentation which yields true values?
\n\nAgain thank you so much to both PIs for being incredibly involved in this discussion, it is absolutely invaluable!
\n\nCheers
", "votes": null, "replies": [] }, { "id": 102377, "author_name": "michaelhansen", "author_url": "", "post_date": "12/21/2015 23:34:15", "content": "Let me see if I can help.
\n\nThere is really nothing reversed about the logic. One way to approach is, as you say, estimate the volume for every time point of the cardiac cycle and report the smallest volume as LVESV and the largest volume as LVEDV. In current clinical practice this is too time consuming because it requires manually tracing the ventricles at lots of time points, so the clinician will eyeball which phases have the largest and smallest volume and then segment only those. It would be much better if the volume was calculated for every phase. It is up to you whether your algorithm estimates which phases are systole and diastole and then calculate the volume in those phases or you calculate the volume for all phases and then report min and max. The latter approach is certainly much simpler conceptually and I would prefer an algorithm that does that since the time course has diagnostic value too, but that is not the topic of the competition.
\n\nGround truth for this competition has been calculated based on hand drawn (by a cardiologist) contours. If we had some instrumentation that yields "true" values, we would just use that.
\n\nHope this helps.
", "votes": null, "replies": [] }, { "id": 103948, "author_name": "udayabhanu", "author_url": "", "post_date": "01/08/2016 04:32:36", "content": "Hi Martin,
\n\nDo you need an RBM network to find smallest and largest cross-sectional areas in slices? Can't you diff successive frames, and then see which ones are non-zeros, and then arrive at cross-sectional areas for slices? Just thinking whether this diffing approach is a viable one or not.
\n\nI am thinking diffing approach is easier and gets ai and bi of simpson's formula also easily.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "102181": "My main concentration so far has been focusing on unsupervised training for identification of the 2 frames (for any arbitrary slice) pertaining to end-systolic and end-diastolic stages of the heart beat cadence.\r\n\r\nWe know that these two stages are defined as the global minimum and maximum of the entire left ventrical volume, thus it seems logical to assume that these global maximums and minimums (for volume) ALSO occur at the global maximum and minimum (for area) of the individual cross-sections for a given slice.\r\n\r\nHowever, one can *imagine* a situation where this might not necessarily be the case - eg. at maximum LV volume, one LV cross section area is slightly smaller than it is, say, slightly after maximum LV volume.\r\n\r\nI have no reason to believe that this is the case, and one would certainly think that maximum volume occurs at the maximum cross-section areas for each slice.\r\n\r\nWhat are peoples thoughts? Any anatomy experts that could shed some light on wether this is likely to be true or not?\r\n\r\nIs anyone else tackling this problem in a similar fashion (first identifying 2 frames from each slice, in which to build a more complex model)? My gut reaction is that it doesn't make sense to use more than 2 frames per slice.\r\n\r\nI have been having success training restricted boltzmann machines with 1 hidden layer and 1 node (essentially what amounts to a logistic regression) on slices. This works out well because we can do it totally unsupervised - we have no labeling pertaining to the cadence. Essentially, we want to train a model to capture a generative distribution for a given slice MRI. The hope is that the single binary output of the model pertains to end-systolic and end-diastolic stages in the slice. \r\n\r\nAs an example, I trained an RBM on slice 9 of study 525 (validation data), by first finding the 128x128 window within the whole image that has the most variance across the frames.\r\n\r\nAfter training on the grey-scale image, we end up with the following weight:\r\n[weight][2]\r\n\r\nWhich when \"off\" will decode to:\r\n[off][3]\r\n\r\nAnd when \"on\" will decode to:\r\n[on][4]\r\n\r\nThis confirms that we have successfully trained an auto-encoding binary feature pertaining to heart beat cadence, and *completely* unsupervised! This is the power of generative distributions.\r\n\r\nNow keep in mind, all we have done here is create a model which is good at assigning a 1 or a 0 to each frame in a way that attempts to capture as much information as possible about the differences between frames. What is ultimately more useful is the dot product of our image vector and the trained weight (essentially, what we would plug in to the sigmoid function). Calculating that dot-product for each frame gives us this beautiful curve:\r\n[curve][5]\r\n\r\nSo it seems our model is indicating that frame 13 and 29 are our likeliest candidates for the end-systolic and end-diastolic stages in the slice. Checking out the whole slice confirms that this is correct:\r\n[slice][6]\r\n\r\nWhadaya think? :)\r\n\r\n [2]: http://imgur.com/zso6tx7\r\n [3]: http://imgur.com/d1gKeOt\r\n [4]: http://imgur.com/hCx8mP8\r\n [5]: http://imgur.com/KlZTXEt\r\n [6]: http://imgur.com/fxbEsTE", "102210": "I may be able to add a brief comment regarding slice synchrony. \r\n\r\nIf I understand your question correctly, you are asking if the timing of systole (smallest volume) and the timing of the smallest cross-sectional area for each slice is always the same?\r\n\r\nThe short answer is no. In a healthy heart, it would be mostly the case. There can be small differences in timing as the electrical depolarization of the muscle cells spreads across the heart, but in a healthy heart, contraction is mostly synchronous across the slices. \r\n\r\nI believe the correct term for the pathological condition where timing is off is \"intraventricular dyssynchrony\", where there is delay or heterogeneity in the timing of contraction in different myocardial segments within the left ventricle. If you google that, you should find more information. \r\n\r\nIntraventricular dyssynchrony would obviously affect cardiac function. Poor synchronization leads (in general) to poor pumping function (i.e., a low ejection fraction).\r\n\r\nHope this helps.", "102234": "[quote=Michael Hansen;102210]\r\n\r\nI may be able to add a brief comment regarding slice synchrony. \r\n\r\nIf I understand your question correctly, you are asking if the timing of systole (smallest volume) and the timing of the smallest cross-sectional area for each slice is always the same?\r\n\r\nThe short answer is no. In a healthy heart, it would be mostly the case. There can be small differences in timing as the electrical depolarization of the muscle cells spreads across the heart, but in a healthy heart, contraction is mostly synchronous across the slices. \r\n\r\nI believe the correct term for the pathological condition where timing is off is \"intraventricular dyssynchrony\", where there is delay or heterogeneity in the timing of contraction in different myocardial segments within the left ventricle. If you google that, you should find more information. \r\n\r\nIntraventricular dyssynchrony would obviously affect cardiac function. Poor synchronization leads (in general) to poor pumping function (i.e., a low ejection fraction).\r\n\r\nHope this helps.\r\n\r\n[/quote]\r\nThis is *exactly* what I was looking for.\r\n\r\nThanks so much Michael!", "102314": "Here is an additional reply from, Andrew Arai:\r\n\r\n\r\n> I’d suggest a slightly different answer on the question about finding end systole and end diastole.\r\n> \r\n> In general, the cine MRI is acquired with a consistent number of\r\n> images across the cardiac cycle. The acquisition is triggered so as to\r\n> coordinate the acquisition with the cardiac cycle. Thus, in general,\r\n> end diastole and end systole should be at a similar time frame across\r\n> the cardiac cycle for all slices. \r\n> \r\n> That assumption is not perfect since there may be heart rate\r\n> variability due to biological variations from breath hold to breath\r\n> hold and may vary slightly as a function of time. So for measurements\r\n> made at the NIH, we pick a column of images at the same end diastolic\r\n> and end systolic time frame even though we may see that for a given\r\n> slice that one frame earlier or later might be better. Thus, the\r\n> reference standard was measured based on a consistent end diastolic\r\n> and end systolic time frame for the whole short axis stack of images.\r\n> \r\n> From a functional perspective, there are disease states where end\r\n> diastole and end systole do not correspond to the minimum and maximum\r\n> cross sectional area for some slices. For example, if a patient has a\r\n> heart attack, most of the slices should have a smaller end systolic\r\n> and larger end diastolic volume. However, slices that are dominated by\r\n> dysfunctional myocardium associated with the heart attack may be\r\n> dyskinetic. Dyskinesis means a segment bulges while other parts of the\r\n> heart are contracting. Therefore, those slices may have a larger cross\r\n> sectional area at end systole than end diastole.\r\n> \r\n> To bring this to a close, I’d personally recommend thinking about the\r\n> finding end systole based on the smallest LV volume and end diastole\r\n> as the largest LV volume as opposed to thinking only slice by slice.\r\n\r\n(Drs. Michael Hansen and Andrew Arai are colleagues at NIH NHLBI, and are the PIs for the Data Science Bowl)", "102368": "Thank you so much to both PIs for the incredibly thoughtful responses alluding to both the practice that the NIH uses, as well as situations whereby frames may not be synchronized (though they generally are), and situations in which maximum and minimum cross-sectional area is not necessarily correlated with end systole and end diastole (though they generally are). This is incredibly helpful information to have for machine-learning practitioners who are not experts in cardiology!\r\n\r\nRegarding Andrew Arai's last statement:\r\n\r\n[quote=Shannon;102314]\r\n> To bring this to a close, I’d personally recommend thinking about the\r\n> finding end systole based on the smallest LV volume and end diastole\r\n> as the largest LV volume as opposed to thinking only slice by slice.\r\n[/quote]\r\n\r\nMy concern is that the logic here seems to be reversed. Andrew is suggesting that we calculate LV volumes, and use this to determine end systole and end diastole. This challenge is essentially about doing the reverse - (ostensibly) determining which frames pertain to end systole and end diastole for a given slice, knowing that these frames necessarily are the best predictors, and using these frames to calculate LV volumes. I'm unsure of what is being suggested here, any clarification would help greatly.\r\n\r\nThis brings to light another question - how have our ground truth volumes been calculated? Have they been calculated by cardiologists based on MRIs, or have they been calculated using some form of instrumentation which yields true values?\r\n\r\nAgain thank you so much to both PIs for being incredibly involved in this discussion, it is absolutely invaluable!\r\n\r\nCheers", "102377": "Let me see if I can help.\r\n\r\nThere is really nothing reversed about the logic. One way to approach is, as you say, estimate the volume for every time point of the cardiac cycle and report the smallest volume as LVESV and the largest volume as LVEDV. In current clinical practice this is too time consuming because it requires manually tracing the ventricles at lots of time points, so the clinician will eyeball which phases have the largest and smallest volume and then segment only those. It would be much better if the volume was calculated for every phase. It is up to you whether your algorithm estimates which phases are systole and diastole and then calculate the volume in those phases or you calculate the volume for all phases and then report min and max. The latter approach is certainly much simpler conceptually and I would prefer an algorithm that does that since the time course has diagnostic value too, but that is not the topic of the competition. \r\n\r\nGround truth for this competition has been calculated based on hand drawn (by a cardiologist) contours. If we had some instrumentation that yields \"true\" values, we would just use that. \r\n\r\nHope this helps.", "103948": "Hi Martin, \r\n\r\nDo you need an RBM network to find smallest and largest cross-sectional areas in slices? Can't you diff successive frames, and then see which ones are non-zeros, and then arrive at cross-sectional areas for slices? Just thinking whether this diffing approach is a viable one or not. \r\n\r\nI am thinking diffing approach is easier and gets ai and bi of simpson's formula also easily." }, "source": "meta" }