Hello kagglers ,
\nFrom what i found out of one month of this competition is that the 3d mri is a 3 dimensional model of the brain right ?
\n
\n As you can see in this 3d mri that the tumor is highlighted .
\n The 3d mri is stored as slices starting from the left of the 3d model to the right . And 2d cnns are being used to classify them right ? . But not all of those slices contain the highlighted tumor right ? So how can 2d cnns work here. 3d cnns have only 585 samples which do you guys think is worth using ?
Hello kagglers ,
\nFrom what i found out of one month of this competition is that the 3d mri is a 3 dimensional model of the brain right ?
\n
\n As you can see in this 3d mri that the tumor is highlighted .
\n The 3d mri is stored as slices starting from the left of the 3d model to the right . And 2d cnns are being used to classify them right ? . But not all of those slices contain the highlighted tumor right ? So how can 2d cnns work here. 3d cnns have only 585 samples which do you guys think is worth using ?
Depends on how you plan to do the task. My score is 0.775 and I have been using 2D CNNs.
", "rawMarkdown": "Depends on how you plan to do the task. My score is 0.775 and I have been using 2D CNNs.", "votes": null }, { "id": "1511137", "postDate": "09/13/2021 06:44:25", "content": "thank you for the help
", "rawMarkdown": "thank you for the help", "votes": null }, { "id": "1511443", "postDate": "09/13/2021 12:30:47", "content": "@Sabin can I know your cv?
", "rawMarkdown": "Sabin can I know your cv?", "votes": null }, { "id": "1511625", "postDate": "09/13/2021 15:09:37", "content": "CV obtained from my way of doing the task doesnot concede with regular CV obtained from models. That means:- I have not made any k-fold cross validation sets and trained models on them and calculated CV. I have a bit of different approach for this.
", "rawMarkdown": "CV obtained from my way of doing the task doesnot concede with regular CV obtained from models. That means:- I have not made any k-fold cross validation sets and trained models on them and calculated CV. I have a bit of different approach for this.", "votes": null }, { "id": "1512125", "postDate": "09/14/2021 02:12:58", "content": "@acharyasabin1997 ,hello, may I ask whether pre-processing on dicom images improve your performance?
\nAs you mentioned you are using 2D CNN model, are you using already developed architectures (for example EfficientNet) or defining your own architecture?
\nI am asking the above questions because I am trying to understand whether the focus should be on pre-processing of images or on developing better models.
", "rawMarkdown": "acharyasabin1997 ,hello, may I ask whether pre-processing on dicom images improve your performance?\n\nAs you mentioned you are using 2D CNN model, are you using already developed architectures (for example EfficientNet) or defining your own architecture?\n\nI am asking the above questions because I am trying to understand whether the focus should be on pre-processing of images or on developing better models.", "votes": null }, { "id": "1512155", "postDate": "09/14/2021 03:21:18", "content": "Actually both data pre-processing and models. It took me a lot of time to figure out making models from existing architecture and changing them and also pre-processing the images.
", "rawMarkdown": "Actually both data pre-processing and models. It took me a lot of time to figure out making models from existing architecture and changing them and also pre-processing the images.", "votes": null }, { "id": "1512166", "postDate": "09/14/2021 03:33:12", "content": "Thank you so much for your reply.
\nIf you don't mind, please could you please elaborate on what you meant by making changes in the existing architecture and what kind of pre-processing might provide better results?
\nFor instance, I am trying resampling, stacking images from different modalities, normalization on the 2D images.
", "rawMarkdown": "Thank you so much for your reply.\n\nIf you don't mind, please could you please elaborate on what you meant by making changes in the existing architecture and what kind of pre-processing might provide better results?\n\nFor instance, I am trying resampling, stacking images from different modalities, normalization on the 2D images.", "votes": null }, { "id": "1512188", "postDate": "09/14/2021 04:25:41", "content": "Preprocessing is pretty same as you have mentioned. Making changes in existing architecture means modifying current network as per our need like fine tuning , adding layers and so on.
", "rawMarkdown": "Preprocessing is pretty same as you have mentioned. Making changes in existing architecture means modifying current network as per our need like fine tuning , adding layers and so on.", "votes": null }, { "id": "1513403", "postDate": "09/15/2021 06:36:54", "content": "@acharyasabin1997 As you mentioned you used 2D network. did you save dataset as 585 jpg or png images?
\nDid you use data augmentation for increase umber of images?
if you used fine tuning model you used transfer learning isn't it?
", "rawMarkdown": "if you used fine tuning model you used transfer learning isn't it?", "votes": null }, { "id": "1513685", "postDate": "09/15/2021 10:31:12", "content": "saved images as png. Used data augmentation and yes, transfer learning.
", "rawMarkdown": "saved images as png. Used data augmentation and yes, transfer learning.", "votes": null }, { "id": "1513931", "postDate": "09/15/2021 14:35:40", "content": "Thanks. I saved 585 images from one modality and used VGG16 model for train this images but i got auc 50 in all epoch and it didn't change. Can you help me with this problem? @acharyasabin1997
", "rawMarkdown": "Thanks. I saved 585 images from one modality and used VGG16 model for train this images but i got auc 50 in all epoch and it didn't change. Can you help me with this problem? @acharyasabin1997", "votes": null }, { "id": "1514126", "postDate": "09/15/2021 18:02:05", "content": "hello,@Mohammadhosein1998 , you might change your learning rate or use a learning rate scheduler to change learning rate after certain number of epoch.
", "rawMarkdown": "hello,@Mohammadhosein1998 , you might change your learning rate or use a learning rate scheduler to change learning rate after certain number of epoch.", "votes": null }, { "id": "1514127", "postDate": "09/15/2021 18:04:26", "content": "@acharyasabin1997 ,may I ask whether you have utilized all four modality or just one modality for training?
\nDuring testing, does png conversion create any error? such as memory error or it worked fine in your case??
", "rawMarkdown": "acharyasabin1997 ,may I ask whether you have utilized all four modality or just one modality for training?\n\nDuring testing, does png conversion create any error? such as memory error or it worked fine in your case??", "votes": null }, { "id": "1514128", "postDate": "09/15/2021 18:04:44", "content": "Thanks but i think my model is over fit and data isn't enough. Did you test 2D model like VGG and get good result?
", "rawMarkdown": "Thanks but i think my model is over fit and data isn't enough. Did you test 2D model like VGG and get good result?", "votes": null }, { "id": "1514152", "postDate": "09/15/2021 18:18:43", "content": "Hello, I tried 2D EfficientNet with all four modality which gives a score of 0.667 but in my case 3D model works better with all modality. Apart from EfficientNet, I worked with DenseNet (3D), self defined 3D model and 2D StrideNet.
", "rawMarkdown": "Hello, I tried 2D EfficientNet with all four modality which gives a score of 0.667 but in my case 3D model works better with all modality. Apart from EfficientNet, I worked with DenseNet (3D), self defined 3D model and 2D StrideNet.", "votes": null }, { "id": "1514340", "postDate": "09/16/2021 01:45:05", "content": "I have used all four modalities for training, png conversion didn't create any error, used a learning rate scheduler to change learning rate after certain number of epoch,tested 2D model like VGG but it didn't work.
", "rawMarkdown": "I have used all four modalities for training, png conversion didn't create any error, used a learning rate scheduler to change learning rate after certain number of epoch,tested 2D model like VGG but it didn't work.", "votes": null }, { "id": "1514372", "postDate": "09/16/2021 02:54:48", "content": "@sabin thank you so much for your reply.
\nDo you believe any particular pre-processing step might improve the performance?
I am just trying to generalize what kind of pre-processing steps in 2D slices might help the model to learn more.
", "rawMarkdown": "sabin thank you so much for your reply. \nDo you believe any particular pre-processing step might improve the performance?\n\nI am just trying to generalize what kind of pre-processing steps in 2D slices might help the model to learn more.", "votes": null }, { "id": "1514414", "postDate": "09/16/2021 04:10:39", "content": "noise reduction, sampling and so on.
", "rawMarkdown": "noise reduction, sampling and so on.", "votes": null }, { "id": "1514752", "postDate": "09/16/2021 12:10:45", "content": "@sabin, could you please elaborate on what you meant by sampling?
", "rawMarkdown": "sabin, could you please elaborate on what you meant by sampling?", "votes": null }, { "id": "1514803", "postDate": "09/16/2021 12:42:48", "content": "take images from each modality that are not blank and are not similar that is to say you could take alternate sequence of images.
", "rawMarkdown": "take images from each modality that are not blank and are not similar that is to say you could take alternate sequence of images.", "votes": null }, { "id": "1515233", "postDate": "09/16/2021 20:44:45", "content": "Hi, thanks for the advice, I am working on thresholding images based on pixel ratios.
\nBtw did you change the orientation of the images? I mean different modalities have different orientation such as axial, saggital or cornonal.
Did you convert your 2D images to single plane ( for instance: all axial) ?
", "rawMarkdown": "Hi, thanks for the advice, I am working on thresholding images based on pixel ratios. \nBtw did you change the orientation of the images? I mean different modalities have different orientation such as axial, saggital or cornonal.\n\nDid you convert your 2D images to single plane ( for instance: all axial) ?", "votes": null }, { "id": "1515346", "postDate": "09/17/2021 03:06:07", "content": "no converison at all
", "rawMarkdown": "no converison at all", "votes": null }, { "id": "1515841", "postDate": "09/17/2021 14:39:30", "content": "Thanks, @sabin for your reply. I really hope you get the highest position in the upcoming days and I hope you will share your approach and kernel after this challenge ends.
", "rawMarkdown": "Thanks, @sabin for your reply. I really hope you get the highest position in the upcoming days and I hope you will share your approach and kernel after this challenge ends.", "votes": null }, { "id": "1516137", "postDate": "09/18/2021 00:47:30", "content": "Yeah, Sure. Thanks mate.
", "rawMarkdown": "Yeah, Sure. Thanks mate.", "votes": null }, { "id": "1516540", "postDate": "09/18/2021 12:30:43", "content": "Do you use external datasets?
", "rawMarkdown": "Do you use external datasets?", "votes": null }, { "id": "1516634", "postDate": "09/18/2021 14:44:00", "content": "No
", "rawMarkdown": "No", "votes": null }, { "id": "1516699", "postDate": "09/18/2021 16:12:09", "content": "I see, thanks for replying.
", "rawMarkdown": "I see, thanks for replying.", "votes": null }, { "id": "1517260", "postDate": "09/19/2021 11:46:35", "content": "https://www.kaggle.com/c/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/271799
", "rawMarkdown": "https://www.kaggle.com/c/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/271799", "votes": null }, { "id": "1527116", "postDate": "09/28/2021 14:25:06", "content": "Please can you tell me were you making predictions per slice or how were you making predictions? I am using 2dnets too but can't solve this issue. Please help.
", "rawMarkdown": "Please can you tell me were you making predictions per slice or how were you making predictions? I am using 2dnets too but can't solve this issue. Please help.", "votes": null }, { "id": "1527675", "postDate": "09/29/2021 02:20:40", "content": "I am making prediction of slices at certain intervals.
", "rawMarkdown": "I am making prediction of slices at certain intervals.", "votes": null }, { "id": "1527746", "postDate": "09/29/2021 04:03:47", "content": "@acharyasabin1997, while inferencing with test data, did you save the dicom file as png in the /kaggle/working directory?
\nIf so,after submission for LB score, does the kernel give any submission error?
", "rawMarkdown": "acharyasabin1997, while inferencing with test data, did you save the dicom file as png in the /kaggle/working directory?\n\nIf so,after submission for LB score, does the kernel give any submission error?", "votes": null }, { "id": "1527753", "postDate": "09/29/2021 04:10:19", "content": "I am not saving dicom file as png while inferencing.
", "rawMarkdown": "I am not saving dicom file as png while inferencing.", "votes": null } ], "comments": [ { "id": 1511126, "author_name": "acharyasabin1997", "author_url": "", "post_date": "09/13/2021 06:32:21", "content": "Depends on how you plan to do the task. My score is 0.775 and I have been using 2D CNNs.
", "votes": null, "replies": [ { "id": 1511137, "author_name": "swaralipibose", "author_url": "", "post_date": "09/13/2021 06:44:25", "content": "thank you for the help
", "votes": null, "replies": [] }, { "id": 1511443, "author_name": "swaralipibose", "author_url": "", "post_date": "09/13/2021 12:30:47", "content": "@Sabin can I know your cv?
", "votes": null, "replies": [] }, { "id": 1511625, "author_name": "acharyasabin1997", "author_url": "", "post_date": "09/13/2021 15:09:37", "content": "CV obtained from my way of doing the task doesnot concede with regular CV obtained from models. That means:- I have not made any k-fold cross validation sets and trained models on them and calculated CV. I have a bit of different approach for this.
", "votes": null, "replies": [] }, { "id": 1512125, "author_name": "wfarzana", "author_url": "", "post_date": "09/14/2021 02:12:58", "content": "@acharyasabin1997 ,hello, may I ask whether pre-processing on dicom images improve your performance?
\nAs you mentioned you are using 2D CNN model, are you using already developed architectures (for example EfficientNet) or defining your own architecture?
\nI am asking the above questions because I am trying to understand whether the focus should be on pre-processing of images or on developing better models.
", "votes": null, "replies": [] }, { "id": 1512155, "author_name": "acharyasabin1997", "author_url": "", "post_date": "09/14/2021 03:21:18", "content": "Actually both data pre-processing and models. It took me a lot of time to figure out making models from existing architecture and changing them and also pre-processing the images.
", "votes": null, "replies": [] }, { "id": 1512166, "author_name": "wfarzana", "author_url": "", "post_date": "09/14/2021 03:33:12", "content": "Thank you so much for your reply.
\nIf you don't mind, please could you please elaborate on what you meant by making changes in the existing architecture and what kind of pre-processing might provide better results?
\nFor instance, I am trying resampling, stacking images from different modalities, normalization on the 2D images.
", "votes": null, "replies": [] }, { "id": 1512188, "author_name": "acharyasabin1997", "author_url": "", "post_date": "09/14/2021 04:25:41", "content": "Preprocessing is pretty same as you have mentioned. Making changes in existing architecture means modifying current network as per our need like fine tuning , adding layers and so on.
", "votes": null, "replies": [] }, { "id": 1513403, "author_name": "mohammadhosein1998", "author_url": "", "post_date": "09/15/2021 06:36:54", "content": "@acharyasabin1997 As you mentioned you used 2D network. did you save dataset as 585 jpg or png images?
\nDid you use data augmentation for increase umber of images?
if you used fine tuning model you used transfer learning isn't it?
", "votes": null, "replies": [] }, { "id": 1513685, "author_name": "acharyasabin1997", "author_url": "", "post_date": "09/15/2021 10:31:12", "content": "saved images as png. Used data augmentation and yes, transfer learning.
", "votes": null, "replies": [] }, { "id": 1513931, "author_name": "mohammadhosein1998", "author_url": "", "post_date": "09/15/2021 14:35:40", "content": "Thanks. I saved 585 images from one modality and used VGG16 model for train this images but i got auc 50 in all epoch and it didn't change. Can you help me with this problem? @acharyasabin1997
", "votes": null, "replies": [] }, { "id": 1514126, "author_name": "wfarzana", "author_url": "", "post_date": "09/15/2021 18:02:05", "content": "hello,@Mohammadhosein1998 , you might change your learning rate or use a learning rate scheduler to change learning rate after certain number of epoch.
", "votes": null, "replies": [] }, { "id": 1514127, "author_name": "wfarzana", "author_url": "", "post_date": "09/15/2021 18:04:26", "content": "@acharyasabin1997 ,may I ask whether you have utilized all four modality or just one modality for training?
\nDuring testing, does png conversion create any error? such as memory error or it worked fine in your case??
", "votes": null, "replies": [] }, { "id": 1514128, "author_name": "mohammadhosein1998", "author_url": "", "post_date": "09/15/2021 18:04:44", "content": "Thanks but i think my model is over fit and data isn't enough. Did you test 2D model like VGG and get good result?
", "votes": null, "replies": [] }, { "id": 1514152, "author_name": "wfarzana", "author_url": "", "post_date": "09/15/2021 18:18:43", "content": "Hello, I tried 2D EfficientNet with all four modality which gives a score of 0.667 but in my case 3D model works better with all modality. Apart from EfficientNet, I worked with DenseNet (3D), self defined 3D model and 2D StrideNet.
", "votes": null, "replies": [] }, { "id": 1514340, "author_name": "acharyasabin1997", "author_url": "", "post_date": "09/16/2021 01:45:05", "content": "I have used all four modalities for training, png conversion didn't create any error, used a learning rate scheduler to change learning rate after certain number of epoch,tested 2D model like VGG but it didn't work.
", "votes": null, "replies": [] }, { "id": 1514372, "author_name": "wfarzana", "author_url": "", "post_date": "09/16/2021 02:54:48", "content": "@sabin thank you so much for your reply.
\nDo you believe any particular pre-processing step might improve the performance?
I am just trying to generalize what kind of pre-processing steps in 2D slices might help the model to learn more.
", "votes": null, "replies": [] }, { "id": 1514414, "author_name": "acharyasabin1997", "author_url": "", "post_date": "09/16/2021 04:10:39", "content": "noise reduction, sampling and so on.
", "votes": null, "replies": [] }, { "id": 1514752, "author_name": "wfarzana", "author_url": "", "post_date": "09/16/2021 12:10:45", "content": "@sabin, could you please elaborate on what you meant by sampling?
", "votes": null, "replies": [] }, { "id": 1514803, "author_name": "acharyasabin1997", "author_url": "", "post_date": "09/16/2021 12:42:48", "content": "take images from each modality that are not blank and are not similar that is to say you could take alternate sequence of images.
", "votes": null, "replies": [] }, { "id": 1515233, "author_name": "wfarzana", "author_url": "", "post_date": "09/16/2021 20:44:45", "content": "Hi, thanks for the advice, I am working on thresholding images based on pixel ratios.
\nBtw did you change the orientation of the images? I mean different modalities have different orientation such as axial, saggital or cornonal.
Did you convert your 2D images to single plane ( for instance: all axial) ?
", "votes": null, "replies": [] }, { "id": 1515346, "author_name": "acharyasabin1997", "author_url": "", "post_date": "09/17/2021 03:06:07", "content": "no converison at all
", "votes": null, "replies": [] }, { "id": 1515841, "author_name": "wfarzana", "author_url": "", "post_date": "09/17/2021 14:39:30", "content": "Thanks, @sabin for your reply. I really hope you get the highest position in the upcoming days and I hope you will share your approach and kernel after this challenge ends.
", "votes": null, "replies": [] }, { "id": 1516137, "author_name": "acharyasabin1997", "author_url": "", "post_date": "09/18/2021 00:47:30", "content": "Yeah, Sure. Thanks mate.
", "votes": null, "replies": [] }, { "id": 1516540, "author_name": "atamazian", "author_url": "", "post_date": "09/18/2021 12:30:43", "content": "Do you use external datasets?
", "votes": null, "replies": [] }, { "id": 1516634, "author_name": "acharyasabin1997", "author_url": "", "post_date": "09/18/2021 14:44:00", "content": "No
", "votes": null, "replies": [] }, { "id": 1516699, "author_name": "atamazian", "author_url": "", "post_date": "09/18/2021 16:12:09", "content": "I see, thanks for replying.
", "votes": null, "replies": [] }, { "id": 1527116, "author_name": "eddwait", "author_url": "", "post_date": "09/28/2021 14:25:06", "content": "Please can you tell me were you making predictions per slice or how were you making predictions? I am using 2dnets too but can't solve this issue. Please help.
", "votes": null, "replies": [] }, { "id": 1527675, "author_name": "acharyasabin1997", "author_url": "", "post_date": "09/29/2021 02:20:40", "content": "I am making prediction of slices at certain intervals.
", "votes": null, "replies": [] }, { "id": 1527746, "author_name": "wfarzana", "author_url": "", "post_date": "09/29/2021 04:03:47", "content": "@acharyasabin1997, while inferencing with test data, did you save the dicom file as png in the /kaggle/working directory?
\nIf so,after submission for LB score, does the kernel give any submission error?
", "votes": null, "replies": [] }, { "id": 1527753, "author_name": "acharyasabin1997", "author_url": "", "post_date": "09/29/2021 04:10:19", "content": "I am not saving dicom file as png while inferencing.
", "votes": null, "replies": [] } ] }, { "id": 1517260, "author_name": "aimind", "author_url": "", "post_date": "09/19/2021 11:46:35", "content": "https://www.kaggle.com/c/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/271799
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1511080": "Hello kagglers ,\nFrom what i found out of one month of this competition is that the 3d mri is a 3 dimensional model of the brain right ?\n\n As you can see in this 3d mri that the tumor is highlighted .\n The 3d mri is stored as slices starting from the left of the 3d model to the right . And 2d cnns are being used to classify them right ? . But not all of those slices contain the highlighted tumor right ? So how can 2d cnns work here. 3d cnns have only 585 samples which do you guys think is worth using ?", "1511126": "Depends on how you plan to do the task. My score is 0.775 and I have been using 2D CNNs.", "1511137": "thank you for the help", "1511443": "Sabin can I know your cv?", "1511625": "CV obtained from my way of doing the task doesnot concede with regular CV obtained from models. That means:- I have not made any k-fold cross validation sets and trained models on them and calculated CV. I have a bit of different approach for this.", "1512125": "acharyasabin1997 ,hello, may I ask whether pre-processing on dicom images improve your performance?\n\nAs you mentioned you are using 2D CNN model, are you using already developed architectures (for example EfficientNet) or defining your own architecture?\n\nI am asking the above questions because I am trying to understand whether the focus should be on pre-processing of images or on developing better models.", "1512155": "Actually both data pre-processing and models. It took me a lot of time to figure out making models from existing architecture and changing them and also pre-processing the images.", "1512166": "Thank you so much for your reply.\n\nIf you don't mind, please could you please elaborate on what you meant by making changes in the existing architecture and what kind of pre-processing might provide better results?\n\nFor instance, I am trying resampling, stacking images from different modalities, normalization on the 2D images.", "1512188": "Preprocessing is pretty same as you have mentioned. Making changes in existing architecture means modifying current network as per our need like fine tuning , adding layers and so on.", "1513403": "acharyasabin1997 As you mentioned you used 2D network. did you save dataset as 585 jpg or png images?\nDid you use data augmentation for increase umber of images?", "1513450": "if you used fine tuning model you used transfer learning isn't it?", "1513685": "saved images as png. Used data augmentation and yes, transfer learning.", "1513931": "Thanks. I saved 585 images from one modality and used VGG16 model for train this images but i got auc 50 in all epoch and it didn't change. Can you help me with this problem? @acharyasabin1997", "1514126": "hello,@Mohammadhosein1998 , you might change your learning rate or use a learning rate scheduler to change learning rate after certain number of epoch.", "1514127": "acharyasabin1997 ,may I ask whether you have utilized all four modality or just one modality for training?\n\nDuring testing, does png conversion create any error? such as memory error or it worked fine in your case??", "1514128": "Thanks but i think my model is over fit and data isn't enough. Did you test 2D model like VGG and get good result?", "1514152": "Hello, I tried 2D EfficientNet with all four modality which gives a score of 0.667 but in my case 3D model works better with all modality. Apart from EfficientNet, I worked with DenseNet (3D), self defined 3D model and 2D StrideNet.", "1514340": "I have used all four modalities for training, png conversion didn't create any error, used a learning rate scheduler to change learning rate after certain number of epoch,tested 2D model like VGG but it didn't work.", "1514372": "sabin thank you so much for your reply. \nDo you believe any particular pre-processing step might improve the performance?\n\nI am just trying to generalize what kind of pre-processing steps in 2D slices might help the model to learn more.", "1514414": "noise reduction, sampling and so on.", "1514752": "sabin, could you please elaborate on what you meant by sampling?", "1514803": "take images from each modality that are not blank and are not similar that is to say you could take alternate sequence of images.", "1515233": "Hi, thanks for the advice, I am working on thresholding images based on pixel ratios. \nBtw did you change the orientation of the images? I mean different modalities have different orientation such as axial, saggital or cornonal.\n\nDid you convert your 2D images to single plane ( for instance: all axial) ?", "1515346": "no converison at all", "1515841": "Thanks, @sabin for your reply. I really hope you get the highest position in the upcoming days and I hope you will share your approach and kernel after this challenge ends.", "1516137": "Yeah, Sure. Thanks mate.", "1516540": "Do you use external datasets?", "1516634": "No", "1516699": "I see, thanks for replying.", "1517260": "https://www.kaggle.com/c/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/271799", "1527116": "Please can you tell me were you making predictions per slice or how were you making predictions? I am using 2dnets too but can't solve this issue. Please help.", "1527675": "I am making prediction of slices at certain intervals.", "1527746": "acharyasabin1997, while inferencing with test data, did you save the dicom file as png in the /kaggle/working directory?\n\nIf so,after submission for LB score, does the kernel give any submission error?", "1527753": "I am not saving dicom file as png while inferencing." }, "source": "meta" }![\"\"](\"https://static-01.hindawi.com/articles/cmmm/volume-2015/450341/figures/450341.fig.001.svgz\")