I have analiced different papers related with brain tumor and M.L., and one of them was using Kaggle datasets in their benchmarks (https://www.mdpi.com/1424-8220/21/6/2222/htm), the conclusion analizing the datasets are that for better performance and 2 clases (tumor - no-tumor), in small dataset you should use Densenet-169 and big dataset, you should ensemble Densenet-169, Inception-V3 and Resnext-50, the part of classification they have used SVM with Kernel RBF which increase the results, so I will present the related datasets and my opinion about how could be used in the context of the competition:
\n-. Small Dataset, https://www.kaggle.com/navoneel/brain-mri-images-for-brain-tumor-detection: 255 images, 155 with tumor and 98 non tumor. These images didn't have he MGMT value, so, for some approximation and use, I think that can be trained for feature extraction, and transfer learning. So, if you have a good model performance, you can use it for segmentation, in the images of the competition, and then, use only the images that were segmented, excluding the others in the training models. Other approaches are welcome!!
\n-. Large Dataset, https://www.kaggle.com/ahmedhamada0/brain-tumor-detection, This database comprises 3000 images, out of which 1500 images contain tumors while the remaining 1500 images are without tumors. My approximation, is that these dataset could be used like the small one, for segmenting the tumors in the image competition, and doing exclusion of images for training.
\n-. BRaTS 2021 Dataset, besides paper's datasets, there is the BRaTS 2021 challenge dataset that I have not tested yet, but seems to be interesting since it contain some data related with MGMT promoter, many thanks to the work of Darien Schettler, for upload https://www.kaggle.com/dschettler8845/brats-2021-task1 and share the notebook untar https://www.kaggle.com/dschettler8845/how-to-load-basic-data-exploration. He has my upvote! And, I think that the approach of Darien, about how to use it, it's the correct, I cited \"You could use the provided segmentation maps (and the provided BraTSIDs) to teach a model how to segment the tumors in the provided competition dataset. This may be useful for determining the MGMT methylation status as it has been known to be correlated with spatial features/attributes (volume, etc.).\"
\nThe other dataset described have 4 clases, https://www.kaggle.com/sartajbhuvaji/brain-tumor-classification-mri, the preliminary results are worst that in the 2-clases experiments, you can change 2 of the clases, non tumor and glioma tumor, by glioma tumor with MGMT and glioma tumor without MGMT, with the clases of the competition, using augmentation and stratifiedshufflesplit (https://scikit-learn.org/stable/modules/cross_validation.html#stratification) with the dataset for balance the datasets.
\nI'll try to comment my work and experiments with these datasets!
", "messages": [ { "id": "1486968", "postDate": "08/23/2021 10:16:58", "content": "I have analiced different papers related with brain tumor and M.L., and one of them was using Kaggle datasets in their benchmarks (https://www.mdpi.com/1424-8220/21/6/2222/htm), the conclusion analizing the datasets are that for better performance and 2 clases (tumor - no-tumor), in small dataset you should use Densenet-169 and big dataset, you should ensemble Densenet-169, Inception-V3 and Resnext-50, the part of classification they have used SVM with Kernel RBF which increase the results, so I will present the related datasets and my opinion about how could be used in the context of the competition:
\n-. Small Dataset, https://www.kaggle.com/navoneel/brain-mri-images-for-brain-tumor-detection: 255 images, 155 with tumor and 98 non tumor. These images didn't have he MGMT value, so, for some approximation and use, I think that can be trained for feature extraction, and transfer learning. So, if you have a good model performance, you can use it for segmentation, in the images of the competition, and then, use only the images that were segmented, excluding the others in the training models. Other approaches are welcome!!
\n-. Large Dataset, https://www.kaggle.com/ahmedhamada0/brain-tumor-detection, This database comprises 3000 images, out of which 1500 images contain tumors while the remaining 1500 images are without tumors. My approximation, is that these dataset could be used like the small one, for segmenting the tumors in the image competition, and doing exclusion of images for training.
\n-. BRaTS 2021 Dataset, besides paper's datasets, there is the BRaTS 2021 challenge dataset that I have not tested yet, but seems to be interesting since it contain some data related with MGMT promoter, many thanks to the work of Darien Schettler, for upload https://www.kaggle.com/dschettler8845/brats-2021-task1 and share the notebook untar https://www.kaggle.com/dschettler8845/how-to-load-basic-data-exploration. He has my upvote! And, I think that the approach of Darien, about how to use it, it's the correct, I cited \"You could use the provided segmentation maps (and the provided BraTSIDs) to teach a model how to segment the tumors in the provided competition dataset. This may be useful for determining the MGMT methylation status as it has been known to be correlated with spatial features/attributes (volume, etc.).\"
\nThe other dataset described have 4 clases, https://www.kaggle.com/sartajbhuvaji/brain-tumor-classification-mri, the preliminary results are worst that in the 2-clases experiments, you can change 2 of the clases, non tumor and glioma tumor, by glioma tumor with MGMT and glioma tumor without MGMT, with the clases of the competition, using augmentation and stratifiedshufflesplit (https://scikit-learn.org/stable/modules/cross_validation.html#stratification) with the dataset for balance the datasets.
\nI'll try to comment my work and experiments with these datasets!
", "rawMarkdown": "## Hi, everybody 🤘!!!\n\nI have analiced different papers related with brain tumor and M.L., and one of them was using Kaggle datasets in their benchmarks (https://www.mdpi.com/1424-8220/21/6/2222/htm), the conclusion analizing the datasets are that for better performance and 2 clases (tumor - no-tumor), in small dataset you should use Densenet-169 and big dataset, you should ensemble Densenet-169, Inception-V3 and Resnext-50, the part of classification they have used SVM with Kernel RBF which increase the results, so I will present the related datasets and my opinion about how could be used in the context of the competition:\n\n -. **Small Dataset**, https://www.kaggle.com/navoneel/brain-mri-images-for-brain-tumor-detection: **255 images, 155 with tumor and 98 non tumor**. These images didn't have he MGMT value, so, for some approximation and use, I think that can be trained for feature extraction, and transfer learning. So, if you have a good model performance, you can use it for segmentation, in the images of the competition, and then, use only the images that were segmented, excluding the others in the training models. Other approaches are welcome!!\n\n\n-. **Large Dataset**, https://www.kaggle.com/ahmedhamada0/brain-tumor-detection, This database **comprises 3000 images, out of which 1500 images contain tumors while the remaining 1500 images are without tumors**. My approximation, is that these dataset could be used like the small one, for segmenting the tumors in the image competition, and doing exclusion of images for training.\n\n\n-. **BRaTS 2021 Dataset**, besides paper's datasets, there is the BRaTS 2021 challenge dataset that I have not tested yet, but seems to be interesting since it contain some data related with MGMT promoter, many thanks to the work of Darien Schettler, for upload https://www.kaggle.com/dschettler8845/brats-2021-task1 and share the notebook untar https://www.kaggle.com/dschettler8845/how-to-load-basic-data-exploration. He has my upvote! And, I think that the approach of Darien, about how to use it, it's the correct, I cited \"*You could use the provided segmentation maps (and the provided BraTSIDs) to teach a model how to segment the tumors in the provided competition dataset. This may be useful for determining the MGMT methylation status as it has been known to be correlated with spatial features/attributes (volume, etc.).*\"\n\n--------------------------------------------------------------------------------------------------\n\nThe other dataset described have 4 clases, https://www.kaggle.com/sartajbhuvaji/brain-tumor-classification-mri, the preliminary results are worst that in the 2-clases experiments, you can change 2 of the clases, **non tumor and glioma tumor, by glioma tumor with MGMT and glioma tumor without MGMT**, with the clases of the competition, using augmentation and stratifiedshufflesplit (https://scikit-learn.org/stable/modules/cross_validation.html#stratification) with the dataset for balance the datasets. \n\nI'll try to comment my work and experiments with these datasets!", "votes": null }, { "id": "1489201", "postDate": "08/24/2021 19:15:13", "content": "@victorfernandezalbor really nice job. these data would certainly help in ways to improve the modelling process. 👍
", "rawMarkdown": "victorfernandezalbor really nice job. these data would certainly help in ways to improve the modelling process. 👍", "votes": null }, { "id": "1491279", "postDate": "08/26/2021 09:19:32", "content": "I have tested some segmentation images from the competition with BRATS Unet 2020, so I have decided to use mi own, with BRATS UNET 2021: https://www.kaggle.com/victorfernandezalbor/brats2021-unet-output-3002151-imgs , lets try if you want!
", "rawMarkdown": "I have tested some segmentation images from the competition with BRATS Unet 2020, so I have decided to use mi own, with BRATS UNET 2021: https://www.kaggle.com/victorfernandezalbor/brats2021-unet-output-3002151-imgs , lets try if you want!", "votes": null }, { "id": "1491623", "postDate": "08/26/2021 13:53:56", "content": "Hi, thanks for the job done.
\nThe link https://www.kaggle.com/victorfernandezalbor/brats2021-unet-output-3002151-imgs is broken. 404 error.
\nPerhaps it is still private.
\nThanks you and happy coding!
Hi, you are right! It was private, try again!
", "rawMarkdown": "Hi, you are right! It was private, try again!", "votes": null }, { "id": "1507556", "postDate": "09/09/2021 10:06:08", "content": "I have tested segmentation with BRATS'21, and It works pretty well, but not works so well for RSNA challenge, I share here the notebook resulting: https://www.kaggle.com/victorfernandezalbor/brats-20-win-nnunet-segment-with-brats-21-rsna
", "rawMarkdown": "I have tested segmentation with BRATS'21, and It works pretty well, but not works so well for RSNA challenge, I share here the notebook resulting: https://www.kaggle.com/victorfernandezalbor/brats-20-win-nnunet-segment-with-brats-21-rsna", "votes": null } ], "comments": [ { "id": 1489201, "author_name": "susant4learning", "author_url": "", "post_date": "08/24/2021 19:15:13", "content": "@victorfernandezalbor really nice job. these data would certainly help in ways to improve the modelling process. 👍
", "votes": null, "replies": [] }, { "id": 1491279, "author_name": "victorfernandezalbor", "author_url": "", "post_date": "08/26/2021 09:19:32", "content": "I have tested some segmentation images from the competition with BRATS Unet 2020, so I have decided to use mi own, with BRATS UNET 2021: https://www.kaggle.com/victorfernandezalbor/brats2021-unet-output-3002151-imgs , lets try if you want!
", "votes": null, "replies": [] }, { "id": 1491623, "author_name": "victorrocco", "author_url": "", "post_date": "08/26/2021 13:53:56", "content": "Hi, thanks for the job done.
\nThe link https://www.kaggle.com/victorfernandezalbor/brats2021-unet-output-3002151-imgs is broken. 404 error.
\nPerhaps it is still private.
\nThanks you and happy coding!
Hi, you are right! It was private, try again!
", "votes": null, "replies": [] } ] }, { "id": 1507556, "author_name": "victorfernandezalbor", "author_url": "", "post_date": "09/09/2021 10:06:08", "content": "I have tested segmentation with BRATS'21, and It works pretty well, but not works so well for RSNA challenge, I share here the notebook resulting: https://www.kaggle.com/victorfernandezalbor/brats-20-win-nnunet-segment-with-brats-21-rsna
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1486968": "## Hi, everybody 🤘!!!\n\nI have analiced different papers related with brain tumor and M.L., and one of them was using Kaggle datasets in their benchmarks (https://www.mdpi.com/1424-8220/21/6/2222/htm), the conclusion analizing the datasets are that for better performance and 2 clases (tumor - no-tumor), in small dataset you should use Densenet-169 and big dataset, you should ensemble Densenet-169, Inception-V3 and Resnext-50, the part of classification they have used SVM with Kernel RBF which increase the results, so I will present the related datasets and my opinion about how could be used in the context of the competition:\n\n -. **Small Dataset**, https://www.kaggle.com/navoneel/brain-mri-images-for-brain-tumor-detection: **255 images, 155 with tumor and 98 non tumor**. These images didn't have he MGMT value, so, for some approximation and use, I think that can be trained for feature extraction, and transfer learning. So, if you have a good model performance, you can use it for segmentation, in the images of the competition, and then, use only the images that were segmented, excluding the others in the training models. Other approaches are welcome!!\n\n\n-. **Large Dataset**, https://www.kaggle.com/ahmedhamada0/brain-tumor-detection, This database **comprises 3000 images, out of which 1500 images contain tumors while the remaining 1500 images are without tumors**. My approximation, is that these dataset could be used like the small one, for segmenting the tumors in the image competition, and doing exclusion of images for training.\n\n\n-. **BRaTS 2021 Dataset**, besides paper's datasets, there is the BRaTS 2021 challenge dataset that I have not tested yet, but seems to be interesting since it contain some data related with MGMT promoter, many thanks to the work of Darien Schettler, for upload https://www.kaggle.com/dschettler8845/brats-2021-task1 and share the notebook untar https://www.kaggle.com/dschettler8845/how-to-load-basic-data-exploration. He has my upvote! And, I think that the approach of Darien, about how to use it, it's the correct, I cited \"*You could use the provided segmentation maps (and the provided BraTSIDs) to teach a model how to segment the tumors in the provided competition dataset. This may be useful for determining the MGMT methylation status as it has been known to be correlated with spatial features/attributes (volume, etc.).*\"\n\n--------------------------------------------------------------------------------------------------\n\nThe other dataset described have 4 clases, https://www.kaggle.com/sartajbhuvaji/brain-tumor-classification-mri, the preliminary results are worst that in the 2-clases experiments, you can change 2 of the clases, **non tumor and glioma tumor, by glioma tumor with MGMT and glioma tumor without MGMT**, with the clases of the competition, using augmentation and stratifiedshufflesplit (https://scikit-learn.org/stable/modules/cross_validation.html#stratification) with the dataset for balance the datasets. \n\nI'll try to comment my work and experiments with these datasets!", "1489201": "victorfernandezalbor really nice job. these data would certainly help in ways to improve the modelling process. 👍", "1491279": "I have tested some segmentation images from the competition with BRATS Unet 2020, so I have decided to use mi own, with BRATS UNET 2021: https://www.kaggle.com/victorfernandezalbor/brats2021-unet-output-3002151-imgs , lets try if you want!", "1491623": "Hi, thanks for the job done.\nThe link https://www.kaggle.com/victorfernandezalbor/brats2021-unet-output-3002151-imgs is broken. 404 error. \nPerhaps it is still private.\nThanks you and happy coding!", "1491727": "Hi, you are right! It was private, try again!", "1507556": "I have tested segmentation with BRATS'21, and It works pretty well, but not works so well for RSNA challenge, I share here the notebook resulting: https://www.kaggle.com/victorfernandezalbor/brats-20-win-nnunet-segment-with-brats-21-rsna" }, "source": "meta" }