I have trained a 2d CNN model, which treats every image independently, but the model failed to learn things from the data. Even the training accuracy increases, the validation accuracy never increases. I tried many methods to avoid overfitting, but I failed. Now I'm thinking to build a 3d CNN model.
\nTo simplify the problem, I am going to build a 3d CNN baseline model that only uses FLAIR images for training. The input is a series of patient's FLAIR images, and the output is the patient's MGMT value. So it is a sequence-to-point model.
\nBut the number of MRI frames is different with each patient and type, how to define the correct input shape for the 3D CNN model?
I have trained a 2d CNN model, which treats every image independently, but the model failed to learn things from the data. Even the training accuracy increases, the validation accuracy never increases. I tried many methods to avoid overfitting, but I failed. Now I'm thinking to build a 3d CNN model.
\nTo simplify the problem, I am going to build a 3d CNN baseline model that only uses FLAIR images for training. The input is a series of patient's FLAIR images, and the output is the patient's MGMT value. So it is a sequence-to-point model.
\nBut the number of MRI frames is different with each patient and type, how to define the correct input shape for the 3D CNN model?
If you look at this https://www.kaggle.com/rluethy/efficientnet3d-with-one-mri-type#Functions-to-load-images
\nyou will notice that getting the total number of images from one scan_id and setting center number of images and then concatenate images by NUM_IMAGES args.
\nso if you set image size 256 and NUM_IMAGES 64 then it's going to be [1, 256, 256, 64]. where 64 is 32 up and down from the center number of images(z-axis). or you can modify this logic as well.
But we can’t assume that every patient in the private test set has that much images for input. For example, if a patient only has one image in FLAIR folder, the model can’t predict such case.
", "rawMarkdown": "But we can’t assume that every patient in the private test set has that much images for input. For example, if a patient only has one image in FLAIR folder, the model can’t predict such case.", "votes": null }, { "id": "1469651", "postDate": "08/13/2021 04:11:36", "content": "@bcghost we do have access to the private test data, it's the test folder and all the images in FLAIR have at least over 20 images. Because of fluid attenuation, it's necessary to take many images.
", "rawMarkdown": "bcghost we *do* have access to the private test data, it's the **test folder** and all the images in FLAIR have at least over 20 images. Because of fluid attenuation, it's necessary to take many images.", "votes": null }, { "id": "1470770", "postDate": "08/13/2021 17:49:16", "content": "Great. Thanks for your information. Where did you find the minimal number of images for each folder? Did you just test it via submission? I only knew the test folder and submission CSV will be replaced when we submit the submission. Did you hard-cord test the number of images by this way?
", "rawMarkdown": "Great. Thanks for your information. Where did you find the minimal number of images for each folder? Did you just test it via submission? I only knew the test folder and submission CSV will be replaced when we submit the submission. Did you hard-cord test the number of images by this way?", "votes": null }, { "id": "1471549", "postDate": "08/14/2021 09:31:42", "content": "As far as I know, we won’t know the minimum number of images for each folder.
", "rawMarkdown": "As far as I know, we won’t know the minimum number of images for each folder.", "votes": null }, { "id": "1471613", "postDate": "08/14/2021 10:38:39", "content": "We don't have access to private test, that's hidden from us.
", "rawMarkdown": "We **don't** have access to private test, that's hidden from us.", "votes": null }, { "id": "1472172", "postDate": "08/14/2021 17:19:11", "content": "So I am considering data argument for images that are less than the slicing number, or let the input shape be [1, height, width, channels]. But I'm not sure if I only put one image into the 3d CNN network, will my 3D network learn the relationship between images?
", "rawMarkdown": "So I am considering data argument for images that are less than the slicing number, or let the input shape be [1, height, width, channels]. But I'm not sure if I only put one image into the 3d CNN network, will my 3D network learn the relationship between images?", "votes": null }, { "id": "1479284", "postDate": "08/18/2021 11:34:04", "content": "364 is the best and max sequence length you can simply crop the center of each image and can let it be 50 x 50
\nand in an overall [None , 364 , 50 , 50 , 3]
", "rawMarkdown": "364 is the best and max sequence length you can simply crop the center of each image and can let it be 50 x 50 \n\nand in an overall [None , 364 , 50 , 50 , 3]", "votes": null } ], "comments": [ { "id": 1468076, "author_name": "jinssaa", "author_url": "", "post_date": "08/12/2021 08:49:27", "content": "If you look at this https://www.kaggle.com/rluethy/efficientnet3d-with-one-mri-type#Functions-to-load-images
\nyou will notice that getting the total number of images from one scan_id and setting center number of images and then concatenate images by NUM_IMAGES args.
\nso if you set image size 256 and NUM_IMAGES 64 then it's going to be [1, 256, 256, 64]. where 64 is 32 up and down from the center number of images(z-axis). or you can modify this logic as well.
But we can’t assume that every patient in the private test set has that much images for input. For example, if a patient only has one image in FLAIR folder, the model can’t predict such case.
", "votes": null, "replies": [] }, { "id": 1469651, "author_name": "aristotle609", "author_url": "", "post_date": "08/13/2021 04:11:36", "content": "@bcghost we do have access to the private test data, it's the test folder and all the images in FLAIR have at least over 20 images. Because of fluid attenuation, it's necessary to take many images.
", "votes": null, "replies": [] }, { "id": 1470770, "author_name": "bcghost", "author_url": "", "post_date": "08/13/2021 17:49:16", "content": "Great. Thanks for your information. Where did you find the minimal number of images for each folder? Did you just test it via submission? I only knew the test folder and submission CSV will be replaced when we submit the submission. Did you hard-cord test the number of images by this way?
", "votes": null, "replies": [] }, { "id": 1471549, "author_name": "hawkeat", "author_url": "", "post_date": "08/14/2021 09:31:42", "content": "As far as I know, we won’t know the minimum number of images for each folder.
", "votes": null, "replies": [] }, { "id": 1471613, "author_name": "mrinath", "author_url": "", "post_date": "08/14/2021 10:38:39", "content": "We don't have access to private test, that's hidden from us.
", "votes": null, "replies": [] }, { "id": 1472172, "author_name": "bcghost", "author_url": "", "post_date": "08/14/2021 17:19:11", "content": "So I am considering data argument for images that are less than the slicing number, or let the input shape be [1, height, width, channels]. But I'm not sure if I only put one image into the 3d CNN network, will my 3D network learn the relationship between images?
", "votes": null, "replies": [] } ] }, { "id": 1479284, "author_name": "swaralipibose", "author_url": "", "post_date": "08/18/2021 11:34:04", "content": "364 is the best and max sequence length you can simply crop the center of each image and can let it be 50 x 50
\nand in an overall [None , 364 , 50 , 50 , 3]
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1467890": "I have trained a 2d CNN model, which treats every image independently, but the model failed to learn things from the data. Even the training accuracy increases, the validation accuracy never increases. I tried many methods to avoid overfitting, but I failed. Now I'm thinking to build a 3d CNN model.\nTo simplify the problem, I am going to build a 3d CNN baseline model that only uses FLAIR images for training. The input is a series of patient's FLAIR images, and the output is the patient's MGMT value. So it is a sequence-to-point model.\nBut the number of MRI frames is different with each patient and type, how to define the correct input shape for the 3D CNN model?", "1468076": "If you look at this [https://www.kaggle.com/rluethy/efficientnet3d-with-one-mri-type#Functions-to-load-images](https://www.kaggle.com/rluethy/efficientnet3d-with-one-mri-type#Functions-to-load-images) \nyou will notice that getting the total number of images from one `scan_id` and setting center number of images and then concatenate images by `NUM_IMAGES` args.\nso if you set `image size` 256 and `NUM_IMAGES` 64 then it's going to be [1, 256, 256, 64]. where 64 is 32 up and down from the center number of images(z-axis). or you can modify this logic as well.", "1468147": "But we can’t assume that every patient in the private test set has that much images for input. For example, if a patient only has one image in FLAIR folder, the model can’t predict such case.", "1469651": "bcghost we *do* have access to the private test data, it's the **test folder** and all the images in FLAIR have at least over 20 images. Because of fluid attenuation, it's necessary to take many images.", "1470770": "Great. Thanks for your information. Where did you find the minimal number of images for each folder? Did you just test it via submission? I only knew the test folder and submission CSV will be replaced when we submit the submission. Did you hard-cord test the number of images by this way?", "1471549": "As far as I know, we won’t know the minimum number of images for each folder.", "1471613": "We **don't** have access to private test, that's hidden from us.", "1472172": "So I am considering data argument for images that are less than the slicing number, or let the input shape be [1, height, width, channels]. But I'm not sure if I only put one image into the 3d CNN network, will my 3D network learn the relationship between images?", "1479284": "364 is the best and max sequence length you can simply crop the center of each image and can let it be 50 x 50 \n\nand in an overall [None , 364 , 50 , 50 , 3]" }, "source": "meta" }