Can the MGMT promoter value be detected visually by experts?
\nI guess that if not, how can we build an efficient model then?
Can the MGMT promoter value be detected visually by experts?
\nI guess that if not, how can we build an efficient model then?
I'm not a radiologist, but I do not think they can distinguish tumors with MGMT from those without just by looking at images.
", "rawMarkdown": "I'm not a radiologist, but I do not think they can distinguish tumors with MGMT from those without just by looking at images.", "votes": null }, { "id": "1512609", "postDate": "09/14/2021 12:06:57", "content": "I am also no expert, but I have worked before on other medical datasets.
\nIn this thread, a radiologist explains how to find glioblastomas (looking for a bright spot on FLAIR sequences). But this does not help us to detect the presence of MGMT promoter methylation. According to this thread, one procedure is genomic DNA extraction (not useful here).
\nAccording to this paper, feature maps such as ADC or rCBF could be used to \"see\" MGMT promoter methylation. However, the problem is that these sequences are generated from DWI as far as I can tell. We have fluid-weighted inversion images (FLAIR), they have diffusion-weighted images (DWI).
\nIn Table 5 of the previous paper, we see that Location + Necrosis gives an AUC of about 0.597. This is consistent with the CV score found by many participants in this competition (CV 0.5 - 0.6).
\nThis means that there is a correlation between tumor location and prediction of MGMT promoter methylation, but tumor location itself is not enough. In a past project, I also found perfusion maps (TMax, ADC, …) to be more useful than \"raw\" data. Hence, if we want better predictions, the same information as ADC/rCBF is necessary.
", "rawMarkdown": "I am also no expert, but I have worked before on other medical datasets.\n\nIn this [thread] (https://www.kaggle.com/c/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/254453), a radiologist explains how to find glioblastomas (looking for a bright spot on FLAIR sequences). But this does not help us to detect the presence of MGMT promoter methylation. According to this [thread](https://www.kaggle.com/c/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/264861), one procedure is genomic DNA extraction (not useful here).\n\nAccording to this [paper](https://bmccancer.biomedcentral.com/articles/10.1186/s12885-018-4114-2), feature maps such as ADC or rCBF could be used to \"see\" MGMT promoter methylation. However, the problem is that these sequences are generated from [DWI](https://radiopaedia.org/articles/apparent-diffusion-coefficient-1) as far as I can tell. We have **fluid**-weighted inversion images (FLAIR), they have **diffusion**-weighted images (DWI).\n\nIn Table 5 of the previous [paper] (https://bmccancer.biomedcentral.com/articles/10.1186/s12885-018-4114-2), we see that Location + Necrosis gives an AUC of about 0.597. This is consistent with the CV score found by many participants in this competition (CV 0.5 - 0.6).\n\nThis means that there is a correlation between tumor location and prediction of MGMT promoter methylation, but tumor location itself is not enough. In a past project, I also found perfusion maps (TMax, ADC, ...) to be more useful than \"raw\" data. Hence, if we want better predictions, the same information as ADC/rCBF is necessary.", "votes": null }, { "id": "1512834", "postDate": "09/14/2021 16:21:12", "content": "I think that there is some features that with deep learning can be extracted in order that you cannot detect visually. And In my opinion it will depends on the deep of the network, just in case that could be some featured related with MGMT. But, also in my opinion these datasets are not big enough for feature extracting and also I think that augmentation will not help, because if there is some feature that identify MGMT promoter and you distorsinate this feature with augmentation, I think that this would be a set-back for the net.
", "rawMarkdown": "I think that there is some features that with deep learning can be extracted in order that you cannot detect visually. And In my opinion it will depends on the deep of the network, just in case that could be some featured related with MGMT. But, also in my opinion these datasets are not big enough for feature extracting and also I think that augmentation will not help, because if there is some feature that identify MGMT promoter and you distorsinate this feature with augmentation, I think that this would be a set-back for the net.", "votes": null }, { "id": "1513393", "postDate": "09/15/2021 06:20:37", "content": "This is extremely helpful
\nBig thanks!
\n\nIn a past project, I also found perfusion maps (TMax, ADC, …) to be more useful than \"raw\" data. Hence, if we want better predictions, the same information as ADC/rCBF is necessary.
\n
Do you think that it is possible to modify competition data to make these information more visible?
", "rawMarkdown": "> In a past project, I also found perfusion maps (TMax, ADC, …) to be more useful than \"raw\" data. Hence, if we want better predictions, the same information as ADC/rCBF is necessary.\n\nDo you think that it is possible to modify competition data to make these information more visible?", "votes": null } ], "comments": [ { "id": 1512107, "author_name": "davidbroberts", "author_url": "", "post_date": "09/14/2021 01:29:16", "content": "I'm not a radiologist, but I do not think they can distinguish tumors with MGMT from those without just by looking at images.
", "votes": null, "replies": [] }, { "id": 1512609, "author_name": "lars123", "author_url": "", "post_date": "09/14/2021 12:06:57", "content": "I am also no expert, but I have worked before on other medical datasets.
\nIn this thread, a radiologist explains how to find glioblastomas (looking for a bright spot on FLAIR sequences). But this does not help us to detect the presence of MGMT promoter methylation. According to this thread, one procedure is genomic DNA extraction (not useful here).
\nAccording to this paper, feature maps such as ADC or rCBF could be used to \"see\" MGMT promoter methylation. However, the problem is that these sequences are generated from DWI as far as I can tell. We have fluid-weighted inversion images (FLAIR), they have diffusion-weighted images (DWI).
\nIn Table 5 of the previous paper, we see that Location + Necrosis gives an AUC of about 0.597. This is consistent with the CV score found by many participants in this competition (CV 0.5 - 0.6).
\nThis means that there is a correlation between tumor location and prediction of MGMT promoter methylation, but tumor location itself is not enough. In a past project, I also found perfusion maps (TMax, ADC, …) to be more useful than \"raw\" data. Hence, if we want better predictions, the same information as ADC/rCBF is necessary.
", "votes": null, "replies": [ { "id": 1513393, "author_name": "noussaons", "author_url": "", "post_date": "09/15/2021 06:20:37", "content": "This is extremely helpful
\nBig thanks!
\n\nIn a past project, I also found perfusion maps (TMax, ADC, …) to be more useful than \"raw\" data. Hence, if we want better predictions, the same information as ADC/rCBF is necessary.
\n
Do you think that it is possible to modify competition data to make these information more visible?
", "votes": null, "replies": [] } ] }, { "id": 1512834, "author_name": "victorfernandezalbor", "author_url": "", "post_date": "09/14/2021 16:21:12", "content": "I think that there is some features that with deep learning can be extracted in order that you cannot detect visually. And In my opinion it will depends on the deep of the network, just in case that could be some featured related with MGMT. But, also in my opinion these datasets are not big enough for feature extracting and also I think that augmentation will not help, because if there is some feature that identify MGMT promoter and you distorsinate this feature with augmentation, I think that this would be a set-back for the net.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1511361": "Can the MGMT promoter value be detected visually by experts?\nI guess that if not, how can we build an efficient model then?", "1512107": "I'm not a radiologist, but I do not think they can distinguish tumors with MGMT from those without just by looking at images.", "1512609": "I am also no expert, but I have worked before on other medical datasets.\n\nIn this [thread] (https://www.kaggle.com/c/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/254453), a radiologist explains how to find glioblastomas (looking for a bright spot on FLAIR sequences). But this does not help us to detect the presence of MGMT promoter methylation. According to this [thread](https://www.kaggle.com/c/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/264861), one procedure is genomic DNA extraction (not useful here).\n\nAccording to this [paper](https://bmccancer.biomedcentral.com/articles/10.1186/s12885-018-4114-2), feature maps such as ADC or rCBF could be used to \"see\" MGMT promoter methylation. However, the problem is that these sequences are generated from [DWI](https://radiopaedia.org/articles/apparent-diffusion-coefficient-1) as far as I can tell. We have **fluid**-weighted inversion images (FLAIR), they have **diffusion**-weighted images (DWI).\n\nIn Table 5 of the previous [paper] (https://bmccancer.biomedcentral.com/articles/10.1186/s12885-018-4114-2), we see that Location + Necrosis gives an AUC of about 0.597. This is consistent with the CV score found by many participants in this competition (CV 0.5 - 0.6).\n\nThis means that there is a correlation between tumor location and prediction of MGMT promoter methylation, but tumor location itself is not enough. In a past project, I also found perfusion maps (TMax, ADC, ...) to be more useful than \"raw\" data. Hence, if we want better predictions, the same information as ADC/rCBF is necessary.", "1512834": "I think that there is some features that with deep learning can be extracted in order that you cannot detect visually. And In my opinion it will depends on the deep of the network, just in case that could be some featured related with MGMT. But, also in my opinion these datasets are not big enough for feature extracting and also I think that augmentation will not help, because if there is some feature that identify MGMT promoter and you distorsinate this feature with augmentation, I think that this would be a set-back for the net.", "1513393": "This is extremely helpful\nBig thanks!", "1518656": "> In a past project, I also found perfusion maps (TMax, ADC, …) to be more useful than \"raw\" data. Hence, if we want better predictions, the same information as ADC/rCBF is necessary.\n\nDo you think that it is possible to modify competition data to make these information more visible?" }, "source": "meta" }