I'm starting to take part in this competition, and my main purpose is to build a 3d solution with MONAI.
\nprogress:
\nI'm starting to take part in this competition, and my main purpose is to build a 3d solution with MONAI.
\nprogress:
\ni spend several weeks on 3d solution. But 3d is (much worse?) worse than 2d, which i cannot understand why.
\nthis is the same case for the previous UW-Madison GI Tract Image Segmentation.
Hi @hengck23 , I'm not sure but according to my experience in UW-Madison GI Tract Image Segmentation, tune a 3d model is much hard than 2d model. I also put weeks to get a reasonable performance.
Update 2: I already trained a segmentation model based on 87 samples that have masks. I randomly picked 80% data for train, and the mean dice score on validation set is around 0.9.
\nAn inference example of 1.2.826.0.1.3680043.5596 is attached.
Now, I will try to make use of this model for the final task. I will share the code later (after I build the full pipeline).
", "rawMarkdown": "Update 2: I already trained a segmentation model based on 87 samples that have masks. I randomly picked 80% data for train, and the mean dice score on validation set is around 0.9.\n\nAn inference example of `1.2.826.0.1.3680043.5596` is attached.\n\nNow, I will try to make use of this model for the final task. I will share the code later (after I build the full pipeline).\n\n", "votes": 1 }, { "id": 1947136, "postDate": "2022-09-20T09:32:40.483Z", "content": "Update 1:
\nI built a notebook which shows how to use MONAI transforms to load dicom images as well as masks.
\nFrom the data section, we know that:
\n\n\nthe NIFTI files consist of segmentation in the sagittal plane, while the DICOM files are in the axial plane
\n
With the help of the ResampleToMatchd transform, the masks will be adjusted to match up with the images, thus we do not need to worry about the orientations.
The link of the notebook is in:
\nhttps://www.kaggle.com/code/yiheng/loading-3d-image-and-mask-with-monai
\nI refer to @itsuki9180 's notebook (Thanks itsuki! ) , and create the animation to show that on my method, images and masks are consistent.
Thanks for sharing! Is there a way to transfer DICOM to nii.gz? Then maybe we can use some of your previous public notebook (training/inference).
", "rawMarkdown": "Thanks for sharing! Is there a way to transfer DICOM to nii.gz? Then maybe we can use some of your previous public notebook (training/inference).", "votes": 1 }, { "id": 1951817, "postDate": "2022-09-23T09:04:56.730Z", "content": "Hi @leonshangguan , yes, I will share the script soon.
", "rawMarkdown": "Hi @leonshangguan , yes, I will share the script soon." } ] }, { "id": 1947143, "postDate": "2022-09-20T09:43:59.110Z", "content": "be careful that some of the dicom images are of different orientation.
\ni find a couple of orientations:
[-1, 0, 0, 0],\n [ 0,-1, 0, 0],\n [ 0, 0, 1, 0],\n [ 0, 0, 0, 1],\n\n\nand\n\n\n [0, -1, 0, 0],\n [ -1, 0, 0, 0],\n [ 0, 0, 1, 0],\n [ 0, 0, 0, 1],\nThanks @hengck23 , glad to see you in this challenge : )
\nLet me update the notebook, and add the Orientationd transform which is about to unify the orientation for all samples.
I had written a converter which reads a dicomdirectory and converts it into NifTi formats, all orientations are taken care natively.
\n# ! pip install nekton\n\nfrom nekton.dcm2nii import Dcm2Nii\nconverter = Dcm2Nii()\nconverted_files = converter.run(dicom_directory='/test_files/CT5N', out_directory='/test_files/CT5N', name='Test')\n# Converted 5 DCM to Nifti; Output stored @ /test_files/CT5N\nprint(converted_files)\n# ['/test_files/CT5N/Test_SmartScore_-_Gated_0.5_sec_20010101000000_5.nii.gz']\nMore details are here https://github.com/deepc-health/nekton
\nAlso if you dont want to convert to Numpy array without converting to NifTi you can use the Monai PydicomReader too it takes care of the orientation too.
an example snippet
\nfrom monai.data.image_reader import PydicomReader\nreader = PydicomReader(swap_ij=True)\n\n# # Read a Series as Numpy array\nx = reader.get_data(path_to_series)[0]\ni spend several weeks on 3d solution. But 3d is (much worse?) worse than 2d, which i cannot understand why.
\nthis is the same case for the previous UW-Madison GI Tract Image Segmentation.
Hi @hengck23 , I'm not sure but according to my experience in UW-Madison GI Tract Image Segmentation, tune a 3d model is much hard than 2d model. I also put weeks to get a reasonable performance.
Update 2: I already trained a segmentation model based on 87 samples that have masks. I randomly picked 80% data for train, and the mean dice score on validation set is around 0.9.
\nAn inference example of 1.2.826.0.1.3680043.5596 is attached.
Now, I will try to make use of this model for the final task. I will share the code later (after I build the full pipeline).
", "votes": 1, "replies": [] }, { "id": 1947136, "author_name": "Yiheng Wang", "author_url": "", "post_date": "2022-09-20T09:32:40.483000", "content": "Update 1:
\nI built a notebook which shows how to use MONAI transforms to load dicom images as well as masks.
\nFrom the data section, we know that:
\n\n\nthe NIFTI files consist of segmentation in the sagittal plane, while the DICOM files are in the axial plane
\n
With the help of the ResampleToMatchd transform, the masks will be adjusted to match up with the images, thus we do not need to worry about the orientations.
The link of the notebook is in:
\nhttps://www.kaggle.com/code/yiheng/loading-3d-image-and-mask-with-monai
\nI refer to @itsuki9180 's notebook (Thanks itsuki! ) , and create the animation to show that on my method, images and masks are consistent.
Thanks for sharing! Is there a way to transfer DICOM to nii.gz? Then maybe we can use some of your previous public notebook (training/inference).
", "votes": 1, "replies": [] }, { "id": 1951817, "author_name": "Yiheng Wang", "author_url": "", "post_date": "2022-09-23T09:04:56.730000", "content": "Hi @leonshangguan , yes, I will share the script soon.
", "votes": 0, "replies": [] } ] }, { "id": 1947143, "author_name": "hengck23", "author_url": "", "post_date": "2022-09-20T09:43:59.110000", "content": "be careful that some of the dicom images are of different orientation.
\ni find a couple of orientations:
[-1, 0, 0, 0],\n [ 0,-1, 0, 0],\n [ 0, 0, 1, 0],\n [ 0, 0, 0, 1],\n\n\nand\n\n\n [0, -1, 0, 0],\n [ -1, 0, 0, 0],\n [ 0, 0, 1, 0],\n [ 0, 0, 0, 1],\nThanks @hengck23 , glad to see you in this challenge : )
\nLet me update the notebook, and add the Orientationd transform which is about to unify the orientation for all samples.
I had written a converter which reads a dicomdirectory and converts it into NifTi formats, all orientations are taken care natively.
\n# ! pip install nekton\n\nfrom nekton.dcm2nii import Dcm2Nii\nconverter = Dcm2Nii()\nconverted_files = converter.run(dicom_directory='/test_files/CT5N', out_directory='/test_files/CT5N', name='Test')\n# Converted 5 DCM to Nifti; Output stored @ /test_files/CT5N\nprint(converted_files)\n# ['/test_files/CT5N/Test_SmartScore_-_Gated_0.5_sec_20010101000000_5.nii.gz']\nMore details are here https://github.com/deepc-health/nekton
\nAlso if you dont want to convert to Numpy array without converting to NifTi you can use the Monai PydicomReader too it takes care of the orientation too.
an example snippet
\nfrom monai.data.image_reader import PydicomReader\nreader = PydicomReader(swap_ij=True)\n\n# # Read a Series as Numpy array\nx = reader.get_data(path_to_series)[0]\n