Based on what I have read, 3D CNNs are not the best way to solve this task, but I would still like to give it a shot. However, I am not completely understanding what preprocessing people are doing to pass it through such a model. First of all, don't we need the same number of slices for each MRI in order to pass it through a CNN? How do we preprocess the image to have standard dimensions for all 4 of the modalities and 585 training instances? Another factor that should be constant before passing through a CNN is the orientation of the image. How can we transform the images provided in the coronal perspective into axial or sagittal (or vice versa)? Is it simply reshaping/reinterpreting the array, or is it more complicated than that?
\nI know I asked many questions here and not all of them are very clear. I would appreciate any feedback and I am happy to clarify any of the questions. Thank you in advance.
", "messages": [ { "id": "1504676", "postDate": "09/06/2021 15:07:00", "content": "Based on what I have read, 3D CNNs are not the best way to solve this task, but I would still like to give it a shot. However, I am not completely understanding what preprocessing people are doing to pass it through such a model. First of all, don't we need the same number of slices for each MRI in order to pass it through a CNN? How do we preprocess the image to have standard dimensions for all 4 of the modalities and 585 training instances? Another factor that should be constant before passing through a CNN is the orientation of the image. How can we transform the images provided in the coronal perspective into axial or sagittal (or vice versa)? Is it simply reshaping/reinterpreting the array, or is it more complicated than that?
\nI know I asked many questions here and not all of them are very clear. I would appreciate any feedback and I am happy to clarify any of the questions. Thank you in advance.
", "rawMarkdown": "Based on what I have read, 3D CNNs are not the best way to solve this task, but I would still like to give it a shot. However, I am not completely understanding what preprocessing people are doing to pass it through such a model. First of all, don't we need the same number of slices for each MRI in order to pass it through a CNN? How do we preprocess the image to have standard dimensions for all 4 of the modalities and 585 training instances? Another factor that should be constant before passing through a CNN is the orientation of the image. How can we transform the images provided in the coronal perspective into axial or sagittal (or vice versa)? Is it simply reshaping/reinterpreting the array, or is it more complicated than that?\n\nI know I asked many questions here and not all of them are very clear. I would appreciate any feedback and I am happy to clarify any of the questions. Thank you in advance.", "votes": null }, { "id": "1504706", "postDate": "09/06/2021 15:35:56", "content": "Check out this notebook to get started: https://www.kaggle.com/rluethy/efficientnet3d-with-one-mri-type Look at the load_dicom_images_3d function, it gets all of the images in the same dimensions
Thank you! This helps a lot!
", "rawMarkdown": "Thank you! This helps a lot!", "votes": null }, { "id": "1507550", "postDate": "09/09/2021 09:57:21", "content": "Just curious, how do you come to a conclusion that 3D CNN is not the best way to solve this task ?
", "rawMarkdown": "Just curious, how do you come to a conclusion that 3D CNN is not the best way to solve this task ?", "votes": null }, { "id": "1508146", "postDate": "09/09/2021 23:24:24", "content": "This discussion has many people stating that their CNNs are not working extremely well.
", "rawMarkdown": "This [discussion ](https://www.kaggle.com/c/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/266173)has many people stating that their CNNs are not working extremely well.", "votes": null } ], "comments": [ { "id": 1504706, "author_name": "returnofsputnik", "author_url": "", "post_date": "09/06/2021 15:35:56", "content": "Check out this notebook to get started: https://www.kaggle.com/rluethy/efficientnet3d-with-one-mri-type Look at the load_dicom_images_3d function, it gets all of the images in the same dimensions
Thank you! This helps a lot!
", "votes": null, "replies": [] } ] }, { "id": 1507550, "author_name": "tungvs", "author_url": "", "post_date": "09/09/2021 09:57:21", "content": "Just curious, how do you come to a conclusion that 3D CNN is not the best way to solve this task ?
", "votes": null, "replies": [ { "id": 1508146, "author_name": "saumandas", "author_url": "", "post_date": "09/09/2021 23:24:24", "content": "This discussion has many people stating that their CNNs are not working extremely well.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1504676": "Based on what I have read, 3D CNNs are not the best way to solve this task, but I would still like to give it a shot. However, I am not completely understanding what preprocessing people are doing to pass it through such a model. First of all, don't we need the same number of slices for each MRI in order to pass it through a CNN? How do we preprocess the image to have standard dimensions for all 4 of the modalities and 585 training instances? Another factor that should be constant before passing through a CNN is the orientation of the image. How can we transform the images provided in the coronal perspective into axial or sagittal (or vice versa)? Is it simply reshaping/reinterpreting the array, or is it more complicated than that?\n\nI know I asked many questions here and not all of them are very clear. I would appreciate any feedback and I am happy to clarify any of the questions. Thank you in advance.", "1504706": "Check out this notebook to get started: https://www.kaggle.com/rluethy/efficientnet3d-with-one-mri-type Look at the `load_dicom_images_3d` function, it gets all of the images in the same dimensions", "1504756": "Thank you! This helps a lot!", "1507550": "Just curious, how do you come to a conclusion that 3D CNN is not the best way to solve this task ?", "1508146": "This [discussion ](https://www.kaggle.com/c/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/266173)has many people stating that their CNNs are not working extremely well." }, "source": "meta" }