![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3163942%2F7f3556e547b3a2cd73958d42f154e43c%2FScreenshot_20250907_185823.png?generation=1757264698009475&alt=media\")
In cases where the aneurysm is located at the vessel junctions and bifurcation points, as aneurysms often do, what is the CoW vessel label assign to that aneurysm? Currently, it looks like even for aneurysms located on junctions and near junctions, there is only one vessel class assigned in terms of location. Is that how the vessel location class is generated?
\nFor example, this aneurysm is located at the origin of superior cerebellar artery (SCA), so arguably it is a \"posterior circulation\" class aneurysm. It touches the basillar tip only slightly. However, the vessel location given by the dataset is \"BA tip\", which is not exactly true. If anything, the aneurysm location given should be SCA posterior aneurysm, and not just labeled as \"BA tip\" as it is not even on the BA tip.
Another example is this aneurysm located at the MCA-ICA-ACA bifurcation, and it is tilted to the right side as part of the M1 segment of MCA. The vessel location given is just \"right ICA\", which is not accurate given the involvement of MCA in this case. ![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3163942%2Fa992e24e42dbc671ace93a84c287064f%2FScreenshot_20250907_185833.png?generation=1757264839485939&alt=media\")
Lastly, it looks like the \"Pcom\" aneurysms are often labeled without the presence of an actual Pcom vessel. These aneurysms that are labelled as \"Pcom aneurysms\" are often actually on the ICA vessel. They are at the ICA where Pcom would typically originate from, but often even when there are no Pcoms the aneurysms are still labeled as \"Pcom\" aneurysm, instead of the more appropriate \"ICA\" aneurysm.
\n(Similarly, the same problem exists for ACA when they are intertwined at A3 to give an accurate left and right label. Or for Acom aneurysms that are at the junction of ACA and Acom.)
\nCan the organizers clarify the aneurysm vessel location annotation protocol, for example, the criteria to determine the vessel label, and if only a single vessel label is given per aneurysm?
\nThanks!
", "messages": [ { "id": 3283473, "postDate": "2025-09-07T17:16:32.993Z", "content": "In cases where the aneurysm is located at the vessel junctions and bifurcation points, as aneurysms often do, what is the CoW vessel label assign to that aneurysm? Currently, it looks like even for aneurysms located on junctions and near junctions, there is only one vessel class assigned in terms of location. Is that how the vessel location class is generated?
\nFor example, this aneurysm is located at the origin of superior cerebellar artery (SCA), so arguably it is a \"posterior circulation\" class aneurysm. It touches the basillar tip only slightly. However, the vessel location given by the dataset is \"BA tip\", which is not exactly true. If anything, the aneurysm location given should be SCA posterior aneurysm, and not just labeled as \"BA tip\" as it is not even on the BA tip.
Another example is this aneurysm located at the MCA-ICA-ACA bifurcation, and it is tilted to the right side as part of the M1 segment of MCA. The vessel location given is just \"right ICA\", which is not accurate given the involvement of MCA in this case.
Lastly, it looks like the \"Pcom\" aneurysms are often labeled without the presence of an actual Pcom vessel. These aneurysms that are labelled as \"Pcom aneurysms\" are often actually on the ICA vessel. They are at the ICA where Pcom would typically originate from, but often even when there are no Pcoms the aneurysms are still labeled as \"Pcom\" aneurysm, instead of the more appropriate \"ICA\" aneurysm.
\n(Similarly, the same problem exists for ACA when they are intertwined at A3 to give an accurate left and right label. Or for Acom aneurysms that are at the junction of ACA and Acom.)
\nCan the organizers clarify the aneurysm vessel location annotation protocol, for example, the criteria to determine the vessel label, and if only a single vessel label is given per aneurysm?
\nThanks!
", "rawMarkdown": "In cases where the aneurysm is located at the vessel junctions and bifurcation points, as aneurysms often do, what is the CoW vessel label assign to that aneurysm? Currently, it looks like even for aneurysms located on junctions and near junctions, there is only one vessel class assigned in terms of location. Is that how the vessel location class is generated?\n\nFor example, this aneurysm is located at the origin of superior cerebellar artery (SCA), so arguably it is a \"posterior circulation\" class aneurysm. It touches the basillar tip only slightly. However, the vessel location given by the dataset is \"BA tip\", which is not exactly true. If anything, the aneurysm location given should be SCA posterior aneurysm, and not just labeled as \"BA tip\" as it is not even on the BA tip. \n\nAnother example is this aneurysm located at the MCA-ICA-ACA bifurcation, and it is tilted to the right side as part of the M1 segment of MCA. The vessel location given is just \"right ICA\", which is not accurate given the involvement of MCA in this case. \n\nLastly, it looks like the \"Pcom\" aneurysms are often labeled without the presence of an actual Pcom vessel. These aneurysms that are labelled as \"Pcom aneurysms\" are often actually on the ICA vessel. They are at the ICA where Pcom would typically originate from, but often even when there are no Pcoms the aneurysms are still labeled as \"Pcom\" aneurysm, instead of the more appropriate \"ICA\" aneurysm.\n\n(Similarly, the same problem exists for ACA when they are intertwined at A3 to give an accurate left and right label. Or for Acom aneurysms that are at the junction of ACA and Acom.)\n\nCan the organizers clarify the aneurysm vessel location annotation protocol, for example, the criteria to determine the vessel label, and if only a single vessel label is given per aneurysm?\n\nThanks!", "votes": 10 }, { "id": 3287124, "postDate": "2025-09-10T22:30:53.773Z", "content": "Happy to provide some insight here. This might be a long one!
\nTo determine aneurysm location, we first asked all of the data contributing sites to review and annotate aneurysm locations and provide these alongside the image data. Next, we recruited a large pool of volunteer neuroradiologist annotators, who subsequently reviewed every single case individually to confirm aneurysm location. Finally, for all cases where there was a discrepancy between the data contributing site and the volunteer annotator (which was a substantial portion!) we manually re-reviewed cases using a much smaller group of subject experts (neuro-interventional radiologists and sub-specialized neuroradiologists) to adjudicate each discrepancy. The final adjudicated scores were considered the reference standard.
\nThe segmentations were generated using a separate process. First we started with a model trained on TopCoW data, used it to pre-segment a subset of our challenge data, heavily modified which vessels and segments were included, and then manually refined each case. We then used this new data to iteratively train new models for vascular segmentation, presegment new data, and then manually refine. In the end, all of the included segmentation data was manually corrected for accuracy. Finally, after we had the final reference standard aneurysm locations as described above, we went back and individually edited the vessel segmentations such that the aneurysms were segmented with the adjudicated parent vessel label.
\nAs you can see we put quite some thought and effort into this. That said, no dataset of this complexity is perfect, so I am sure there are some errors. We can only hope they are minor! Hope this answers your question.
\nAlso worth noting (as you likely know) that sometimes its not possible to be definitive about the parent vessel using non-invasive imaging. In these cases, we just did our best \"educated guess\".
", "rawMarkdown": "Happy to provide some insight here. This might be a long one!\n\nTo determine aneurysm location, we first asked all of the data contributing sites to review and annotate aneurysm locations and provide these alongside the image data. Next, we recruited a large pool of volunteer neuroradiologist annotators, who subsequently reviewed every single case individually to confirm aneurysm location. Finally, for all cases where there was a discrepancy between the data contributing site and the volunteer annotator (which was a substantial portion!) we manually re-reviewed cases using a much smaller group of subject experts (neuro-interventional radiologists and sub-specialized neuroradiologists) to adjudicate each discrepancy. The final adjudicated scores were considered the reference standard.\n\nThe segmentations were generated using a separate process. First we started with a model trained on TopCoW data, used it to pre-segment a subset of our challenge data, heavily modified which vessels and segments were included, and then manually refined each case. We then used this new data to iteratively train new models for vascular segmentation, presegment new data, and then manually refine. In the end, all of the included segmentation data was manually corrected for accuracy. Finally, after we had the final reference standard aneurysm locations as described above, we went back and individually edited the vessel segmentations such that the aneurysms were segmented with the adjudicated parent vessel label.\n\nAs you can see we put quite some thought and effort into this. That said, no dataset of this complexity is perfect, so I am sure there are some errors. We can only hope they are minor! Hope this answers your question. \n\nAlso worth noting (as you likely know) that sometimes its not possible to be definitive about the parent vessel using non-invasive imaging. In these cases, we just did our best \"educated guess\".", "votes": 6, "replies": [ { "id": 3287171, "postDate": "2025-09-11T02:50:08.993Z", "content": "Thank you for your great work — this appears to be a massive and impressive effort!
\nYou mentioned using Topcow (a publicly available and free tool) to generate the initial vessel labels. I noticed that its license prohibits commercial use. Does this comply with Kaggle's rules regarding the use of external data?
\nAdditionally, would it be permissible for us to include Topcow data in our training dataset for training vessel segmentation models?
\nThank you very much for your clarification!
", "rawMarkdown": "Thank you for your great work — this appears to be a massive and impressive effort!\n\nYou mentioned using Topcow (a publicly available and free tool) to generate the initial vessel labels. I noticed that its license prohibits commercial use. Does this comply with Kaggle's rules regarding the use of external data?\n\nAdditionally, would it be permissible for us to include Topcow data in our training dataset for training vessel segmentation models?\n\nThank you very much for your clarification!", "votes": 1, "replies": [ { "id": 3288057, "postDate": "2025-09-12T23:09:51.600Z", "content": "Prohibiting commercial use in the data license means you cant use this data to create and sell a product. It doesn't have anything to do with how RSNA created this dataset.
\nYou are free to use the Topcow data since it is publicly available external dataset, but don't think it would be very useful as we made significant changes to the labeling schema, so adding it as more training data probably wouldn't work very well.
I see, thanks a lot
", "rawMarkdown": "I see, thanks a lot" } ] } ] } ], "comments": [ { "id": 3287124, "author_name": "Evan Calabrese", "author_url": "", "post_date": "2025-09-10T22:30:53.773000", "content": "Happy to provide some insight here. This might be a long one!
\nTo determine aneurysm location, we first asked all of the data contributing sites to review and annotate aneurysm locations and provide these alongside the image data. Next, we recruited a large pool of volunteer neuroradiologist annotators, who subsequently reviewed every single case individually to confirm aneurysm location. Finally, for all cases where there was a discrepancy between the data contributing site and the volunteer annotator (which was a substantial portion!) we manually re-reviewed cases using a much smaller group of subject experts (neuro-interventional radiologists and sub-specialized neuroradiologists) to adjudicate each discrepancy. The final adjudicated scores were considered the reference standard.
\nThe segmentations were generated using a separate process. First we started with a model trained on TopCoW data, used it to pre-segment a subset of our challenge data, heavily modified which vessels and segments were included, and then manually refined each case. We then used this new data to iteratively train new models for vascular segmentation, presegment new data, and then manually refine. In the end, all of the included segmentation data was manually corrected for accuracy. Finally, after we had the final reference standard aneurysm locations as described above, we went back and individually edited the vessel segmentations such that the aneurysms were segmented with the adjudicated parent vessel label.
\nAs you can see we put quite some thought and effort into this. That said, no dataset of this complexity is perfect, so I am sure there are some errors. We can only hope they are minor! Hope this answers your question.
\nAlso worth noting (as you likely know) that sometimes its not possible to be definitive about the parent vessel using non-invasive imaging. In these cases, we just did our best \"educated guess\".
", "votes": 6, "replies": [ { "id": 3287171, "author_name": "Shuolin Liu", "author_url": "", "post_date": "2025-09-11T02:50:08.993000", "content": "Thank you for your great work — this appears to be a massive and impressive effort!
\nYou mentioned using Topcow (a publicly available and free tool) to generate the initial vessel labels. I noticed that its license prohibits commercial use. Does this comply with Kaggle's rules regarding the use of external data?
\nAdditionally, would it be permissible for us to include Topcow data in our training dataset for training vessel segmentation models?
\nThank you very much for your clarification!
", "votes": 1, "replies": [ { "id": 3288057, "author_name": "JeffRudie", "author_url": "", "post_date": "2025-09-12T23:09:51.600000", "content": "Prohibiting commercial use in the data license means you cant use this data to create and sell a product. It doesn't have anything to do with how RSNA created this dataset.
\nYou are free to use the Topcow data since it is publicly available external dataset, but don't think it would be very useful as we made significant changes to the labeling schema, so adding it as more training data probably wouldn't work very well.
I see, thanks a lot
", "votes": 0, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "3283473": "In cases where the aneurysm is located at the vessel junctions and bifurcation points, as aneurysms often do, what is the CoW vessel label assign to that aneurysm? Currently, it looks like even for aneurysms located on junctions and near junctions, there is only one vessel class assigned in terms of location. Is that how the vessel location class is generated?\n\nFor example, this aneurysm is located at the origin of superior cerebellar artery (SCA), so arguably it is a \"posterior circulation\" class aneurysm. It touches the basillar tip only slightly. However, the vessel location given by the dataset is \"BA tip\", which is not exactly true. If anything, the aneurysm location given should be SCA posterior aneurysm, and not just labeled as \"BA tip\" as it is not even on the BA tip. \n\nAnother example is this aneurysm located at the MCA-ICA-ACA bifurcation, and it is tilted to the right side as part of the M1 segment of MCA. The vessel location given is just \"right ICA\", which is not accurate given the involvement of MCA in this case. \n\nLastly, it looks like the \"Pcom\" aneurysms are often labeled without the presence of an actual Pcom vessel. These aneurysms that are labelled as \"Pcom aneurysms\" are often actually on the ICA vessel. They are at the ICA where Pcom would typically originate from, but often even when there are no Pcoms the aneurysms are still labeled as \"Pcom\" aneurysm, instead of the more appropriate \"ICA\" aneurysm.\n\n(Similarly, the same problem exists for ACA when they are intertwined at A3 to give an accurate left and right label. Or for Acom aneurysms that are at the junction of ACA and Acom.)\n\nCan the organizers clarify the aneurysm vessel location annotation protocol, for example, the criteria to determine the vessel label, and if only a single vessel label is given per aneurysm?\n\nThanks!", "3287124": "Happy to provide some insight here. This might be a long one!\n\nTo determine aneurysm location, we first asked all of the data contributing sites to review and annotate aneurysm locations and provide these alongside the image data. Next, we recruited a large pool of volunteer neuroradiologist annotators, who subsequently reviewed every single case individually to confirm aneurysm location. Finally, for all cases where there was a discrepancy between the data contributing site and the volunteer annotator (which was a substantial portion!) we manually re-reviewed cases using a much smaller group of subject experts (neuro-interventional radiologists and sub-specialized neuroradiologists) to adjudicate each discrepancy. The final adjudicated scores were considered the reference standard.\n\nThe segmentations were generated using a separate process. First we started with a model trained on TopCoW data, used it to pre-segment a subset of our challenge data, heavily modified which vessels and segments were included, and then manually refined each case. We then used this new data to iteratively train new models for vascular segmentation, presegment new data, and then manually refine. In the end, all of the included segmentation data was manually corrected for accuracy. Finally, after we had the final reference standard aneurysm locations as described above, we went back and individually edited the vessel segmentations such that the aneurysms were segmented with the adjudicated parent vessel label.\n\nAs you can see we put quite some thought and effort into this. That said, no dataset of this complexity is perfect, so I am sure there are some errors. We can only hope they are minor! Hope this answers your question. \n\nAlso worth noting (as you likely know) that sometimes its not possible to be definitive about the parent vessel using non-invasive imaging. In these cases, we just did our best \"educated guess\"." } }