I have some kind of intuition (and also some quantitive analysis behind) that there may be clear differences in reliability of the MRI-data regarding to the MGMT-analysis. In the graph https://github.com/mih-ai/tukimateriaalia/blob/main/suorituskykykuva.png is shown some (one statistical sample) of the error levels in MGMT analysis in light of measurement type and measurement time (inside the folder).
\nConclusions, referring to the picture in the link above, seem to be that:
\n-FLAIR produces (as independent…) the most reliable results, the T2w is the next, but overall, all the measurement types seem to be useful at least in some cases if intelligently utilized. Means; all data may include useful information in training, but especially when making inference, weighting the analysis based on the \"reliability order\" may be useful to obtain better results.
\n-The measurements taken near the middle of the image folder are most useful. Means; all data may include useful information, but weighting the data locating \"in time domain\" near the center is most reliable.
\nAnyone who would get advantage about the thoughts above, please utilize the findings freely, and happy to hear other's similar kind on findings. Do you colleguages have similar kind of findings or evidences?
\nLähetän nämä havainnot kilpailusääntöjen mukaan julkisesti mutta samalla pohdintana erityisesti suomalaisille tekoälystä kiinnostuneille että heitetäänpä tekoälykollegat \"läppää\" tästä aiheesta täällä suomeksi tai englanniksi ja ollaan aktiivisia, aihe on kiintoisa. Onkohan meitä tässä kisassa mukana leaderboardilla asti muita suomalaisia tällä hetkellä kuin allekirjoittanut ja @qitvision ja kumppanit?
", "messages": [ { "id": "1534712", "postDate": "10/05/2021 07:21:02", "content": "I have some kind of intuition (and also some quantitive analysis behind) that there may be clear differences in reliability of the MRI-data regarding to the MGMT-analysis. In the graph https://github.com/mih-ai/tukimateriaalia/blob/main/suorituskykykuva.png is shown some (one statistical sample) of the error levels in MGMT analysis in light of measurement type and measurement time (inside the folder).
\nConclusions, referring to the picture in the link above, seem to be that:
\n-FLAIR produces (as independent…) the most reliable results, the T2w is the next, but overall, all the measurement types seem to be useful at least in some cases if intelligently utilized. Means; all data may include useful information in training, but especially when making inference, weighting the analysis based on the \"reliability order\" may be useful to obtain better results.
\n-The measurements taken near the middle of the image folder are most useful. Means; all data may include useful information, but weighting the data locating \"in time domain\" near the center is most reliable.
\nAnyone who would get advantage about the thoughts above, please utilize the findings freely, and happy to hear other's similar kind on findings. Do you colleguages have similar kind of findings or evidences?
\nLähetän nämä havainnot kilpailusääntöjen mukaan julkisesti mutta samalla pohdintana erityisesti suomalaisille tekoälystä kiinnostuneille että heitetäänpä tekoälykollegat \"läppää\" tästä aiheesta täällä suomeksi tai englanniksi ja ollaan aktiivisia, aihe on kiintoisa. Onkohan meitä tässä kisassa mukana leaderboardilla asti muita suomalaisia tällä hetkellä kuin allekirjoittanut ja @qitvision ja kumppanit?
", "rawMarkdown": "I have some kind of intuition (and also some quantitive analysis behind) that there may be clear differences in reliability of the MRI-data regarding to the MGMT-analysis. In the graph https://github.com/mih-ai/tukimateriaalia/blob/main/suorituskykykuva.png is shown some (one statistical sample) of the error levels in MGMT analysis in light of measurement type and measurement time (inside the folder). \n\nConclusions, referring to the picture in the link above, seem to be that:\n\n-FLAIR produces (as independent...) the most reliable results, the T2w is the next, but overall, all the measurement types seem to be useful at least in some cases if intelligently utilized. Means; all data may include useful information in training, but especially when making inference, weighting the analysis based on the \"reliability order\" may be useful to obtain better results. \n\n-The measurements taken near the middle of the image folder are most useful. Means; all data may include useful information, but weighting the data locating \"in time domain\" near the center is most reliable. \n\nAnyone who would get advantage about the thoughts above, please utilize the findings freely, and happy to hear other's similar kind on findings. Do you colleguages have similar kind of findings or evidences?\n\nLähetän nämä havainnot kilpailusääntöjen mukaan julkisesti mutta samalla pohdintana erityisesti suomalaisille tekoälystä kiinnostuneille että heitetäänpä tekoälykollegat \"läppää\" tästä aiheesta täällä suomeksi tai englanniksi ja ollaan aktiivisia, aihe on kiintoisa. Onkohan meitä tässä kisassa mukana leaderboardilla asti muita suomalaisia tällä hetkellä kuin allekirjoittanut ja @qitvision ja kumppanit?", "votes": null }, { "id": "1534999", "postDate": "10/05/2021 12:09:38", "content": "I have a similar impression that FLAIR and T2W seem like the best modalities for the task. T2W has been working according to literature, and FLAIR highlights possibly relevant features (non-free-flowing water and Edema). More on these threads:
\nhttps://www.kaggle.com/c/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/273671
\nhttps://www.kaggle.com/c/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/273525
\nRather than using the Dicom stacks as time-series images, I find it's easier to work with the MRI data after registering to normalized voxels like in this notebook. Registering doesn't solve the issue of where to pick a slice for a 2D CNN, though.
", "rawMarkdown": "I have a similar impression that FLAIR and T2W seem like the best modalities for the task. T2W has been working according to literature, and FLAIR highlights possibly relevant features (non-free-flowing water and Edema). More on these threads:\n\nhttps://www.kaggle.com/c/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/273671\n\nhttps://www.kaggle.com/c/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/273525\n\nRather than using the Dicom stacks as time-series images, I find it's easier to work with the MRI data after registering to normalized voxels [like in this notebook](https://www.kaggle.com/ren4yu/normalized-voxels-align-planes-and-crop). Registering doesn't solve the issue of where to pick a slice for a 2D CNN, though.", "votes": null }, { "id": "1549580", "postDate": "10/19/2021 04:28:24", "content": "Congrats @qitvision & rähmä.ai 🙂
", "rawMarkdown": "Congrats @qitvision & rähmä.ai 🙂", "votes": null } ], "comments": [ { "id": 1534999, "author_name": "qitvision", "author_url": "", "post_date": "10/05/2021 12:09:38", "content": "I have a similar impression that FLAIR and T2W seem like the best modalities for the task. T2W has been working according to literature, and FLAIR highlights possibly relevant features (non-free-flowing water and Edema). More on these threads:
\nhttps://www.kaggle.com/c/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/273671
\nhttps://www.kaggle.com/c/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/273525
\nRather than using the Dicom stacks as time-series images, I find it's easier to work with the MRI data after registering to normalized voxels like in this notebook. Registering doesn't solve the issue of where to pick a slice for a 2D CNN, though.
", "votes": null, "replies": [] }, { "id": 1549580, "author_name": "experienceinai", "author_url": "", "post_date": "10/19/2021 04:28:24", "content": "Congrats @qitvision & rähmä.ai 🙂
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1534712": "I have some kind of intuition (and also some quantitive analysis behind) that there may be clear differences in reliability of the MRI-data regarding to the MGMT-analysis. In the graph https://github.com/mih-ai/tukimateriaalia/blob/main/suorituskykykuva.png is shown some (one statistical sample) of the error levels in MGMT analysis in light of measurement type and measurement time (inside the folder). \n\nConclusions, referring to the picture in the link above, seem to be that:\n\n-FLAIR produces (as independent...) the most reliable results, the T2w is the next, but overall, all the measurement types seem to be useful at least in some cases if intelligently utilized. Means; all data may include useful information in training, but especially when making inference, weighting the analysis based on the \"reliability order\" may be useful to obtain better results. \n\n-The measurements taken near the middle of the image folder are most useful. Means; all data may include useful information, but weighting the data locating \"in time domain\" near the center is most reliable. \n\nAnyone who would get advantage about the thoughts above, please utilize the findings freely, and happy to hear other's similar kind on findings. Do you colleguages have similar kind of findings or evidences?\n\nLähetän nämä havainnot kilpailusääntöjen mukaan julkisesti mutta samalla pohdintana erityisesti suomalaisille tekoälystä kiinnostuneille että heitetäänpä tekoälykollegat \"läppää\" tästä aiheesta täällä suomeksi tai englanniksi ja ollaan aktiivisia, aihe on kiintoisa. Onkohan meitä tässä kisassa mukana leaderboardilla asti muita suomalaisia tällä hetkellä kuin allekirjoittanut ja @qitvision ja kumppanit?", "1534999": "I have a similar impression that FLAIR and T2W seem like the best modalities for the task. T2W has been working according to literature, and FLAIR highlights possibly relevant features (non-free-flowing water and Edema). More on these threads:\n\nhttps://www.kaggle.com/c/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/273671\n\nhttps://www.kaggle.com/c/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/273525\n\nRather than using the Dicom stacks as time-series images, I find it's easier to work with the MRI data after registering to normalized voxels [like in this notebook](https://www.kaggle.com/ren4yu/normalized-voxels-align-planes-and-crop). Registering doesn't solve the issue of where to pick a slice for a 2D CNN, though.", "1549580": "Congrats @qitvision & rähmä.ai 🙂" }, "source": "meta" }