I have noticed that the training set is provided in the form of slices, whereas medical CT images are typically presented as 3D images.
\nCT images can exhibit various characteristics depending on the scanning machine, such as voxel spacing. Voxel spacing refers to the distance between neighboring voxels in three spatial dimensions (x, y, and z axes) within CT images. It represents the actual size of a pixel in a CT image in physical space. After examining a few individual slices, I discovered that their voxel spacing differed, implying that resampling the CT image might be necessary. However, the dataset only provides slices and lacks continuity, which poses challenges for data processing.
\nIn terms of the issues that inconsistent voxel spacing can cause in deep learning:
\nAdjusting the spacing of a CT image can impact its overall pixel value and contextual information. Increasing spacing reduces the overall pixel value while enhancing contextual information, whereas decreasing spacing increases the overall pixel value while reducing contextual information. Therefore, striking the right balance between the amount of contextual information in the network's patch size and the preservation of image details is crucial for achieving optimal performance.
Deep learning models typically perform convolution operations on localized regions of the input data and share weights to extract features. Inconsistencies in voxel spacing among input images can disrupt weight sharing, potentially affecting the model's performance.
Introducing CT images with varying voxel spacing to the model may result in challenges during feature extraction and matching due to inconsistent correspondence between pixels. This, in turn, can impact the model's performance.
Maybe I'm misunderstanding the task? It's not a 3d image mission, just a 2d?
", "messages": [ { "id": 2362356, "postDate": "2023-07-28T03:40:09.933Z", "content": "I have noticed that the training set is provided in the form of slices, whereas medical CT images are typically presented as 3D images.
\nCT images can exhibit various characteristics depending on the scanning machine, such as voxel spacing. Voxel spacing refers to the distance between neighboring voxels in three spatial dimensions (x, y, and z axes) within CT images. It represents the actual size of a pixel in a CT image in physical space. After examining a few individual slices, I discovered that their voxel spacing differed, implying that resampling the CT image might be necessary. However, the dataset only provides slices and lacks continuity, which poses challenges for data processing.
\nIn terms of the issues that inconsistent voxel spacing can cause in deep learning:
\nAdjusting the spacing of a CT image can impact its overall pixel value and contextual information. Increasing spacing reduces the overall pixel value while enhancing contextual information, whereas decreasing spacing increases the overall pixel value while reducing contextual information. Therefore, striking the right balance between the amount of contextual information in the network's patch size and the preservation of image details is crucial for achieving optimal performance.
Deep learning models typically perform convolution operations on localized regions of the input data and share weights to extract features. Inconsistencies in voxel spacing among input images can disrupt weight sharing, potentially affecting the model's performance.
Introducing CT images with varying voxel spacing to the model may result in challenges during feature extraction and matching due to inconsistent correspondence between pixels. This, in turn, can impact the model's performance.
Maybe I'm misunderstanding the task? It's not a 3d image mission, just a 2d?
", "rawMarkdown": "I have noticed that the training set is provided in the form of slices, whereas medical CT images are typically presented as 3D images.\n\nCT images can exhibit various characteristics depending on the scanning machine, such as voxel spacing. Voxel spacing refers to the distance between neighboring voxels in three spatial dimensions (x, y, and z axes) within CT images. It represents the actual size of a pixel in a CT image in physical space. After examining a few individual slices, I discovered that their voxel spacing differed, implying that resampling the CT image might be necessary. However, the dataset only provides slices and lacks continuity, which poses challenges for data processing.\n\nIn terms of the issues that inconsistent voxel spacing can cause in deep learning:\n1. Adjusting the spacing of a CT image can impact its overall pixel value and contextual information. Increasing spacing reduces the overall pixel value while enhancing contextual information, whereas decreasing spacing increases the overall pixel value while reducing contextual information. Therefore, striking the right balance between the amount of contextual information in the network's patch size and the preservation of image details is crucial for achieving optimal performance.\n\n2. Deep learning models typically perform convolution operations on localized regions of the input data and share weights to extract features. Inconsistencies in voxel spacing among input images can disrupt weight sharing, potentially affecting the model's performance.\n\n3. Introducing CT images with varying voxel spacing to the model may result in challenges during feature extraction and matching due to inconsistent correspondence between pixels. This, in turn, can impact the model's performance.\n\n\n\nMaybe I'm misunderstanding the task? It's not a 3d image mission, just a 2d?\n\n\n\n", "votes": 6 }, { "id": 2364742, "postDate": "2023-07-29T15:09:26.750Z", "content": "It is common to see varying voxel dimensions and slice \"thickness/distance\" in CT. Machines are different and scanning protocols are different, depending on the anatomy/pathology in question. All CT modalities acquire a 3D 'volume' of data, then reconstruct the data into 2D slices in various \"planes\" .. or orientations, which are the images we can see. Those images can then be reconstructed into other planes or volumes depending on the clinical need.
\nIdeally, we'd scan every patient at the highest resolution possible. But for CT, ionizing radiation makes that a bad idea. There's always a trade-off between radiation dosage and image detail.
\nChanging voxel sizes, or pixel dimension/spacing data for resampling will distort the anatomy .. which is probably not a good idea.
", "rawMarkdown": "It is common to see varying voxel dimensions and slice \"thickness/distance\" in CT. Machines are different and scanning protocols are different, depending on the anatomy/pathology in question. All CT modalities acquire a 3D 'volume' of data, then reconstruct the data into 2D slices in various \"planes\" .. or orientations, which are the images we can see. Those images can then be reconstructed into other planes or volumes depending on the clinical need.\n\nIdeally, we'd scan every patient at the highest resolution possible. But for CT, ionizing radiation makes that a bad idea. There's always a trade-off between radiation dosage and image detail.\n\nChanging voxel sizes, or pixel dimension/spacing data for resampling will distort the anatomy .. which is probably not a good idea.", "votes": 1 }, { "id": 2363875, "postDate": "2023-07-29T02:45:31.683Z", "content": "It is written (in \"Data\") that \"…scans from dozens of different CT machines have been reprocessed to use the run length encoded lossless compression format but retain other differences such as the number of bits per pixel, pixel range, and pixel representation\" Probably, we need to adapt to this.
", "rawMarkdown": "It is written (in \"Data\") that \"...scans from dozens of different CT machines have been reprocessed to use the run length encoded lossless compression format but retain other differences such as the number of bits per pixel, pixel range, and pixel representation\" Probably, we need to adapt to this." } ], "comments": [ { "id": 2364742, "author_name": "David Roberts", "author_url": "", "post_date": "2023-07-29T15:09:26.750000", "content": "It is common to see varying voxel dimensions and slice \"thickness/distance\" in CT. Machines are different and scanning protocols are different, depending on the anatomy/pathology in question. All CT modalities acquire a 3D 'volume' of data, then reconstruct the data into 2D slices in various \"planes\" .. or orientations, which are the images we can see. Those images can then be reconstructed into other planes or volumes depending on the clinical need.
\nIdeally, we'd scan every patient at the highest resolution possible. But for CT, ionizing radiation makes that a bad idea. There's always a trade-off between radiation dosage and image detail.
\nChanging voxel sizes, or pixel dimension/spacing data for resampling will distort the anatomy .. which is probably not a good idea.
", "votes": 1, "replies": [] }, { "id": 2363875, "author_name": "Pavel Orlov", "author_url": "", "post_date": "2023-07-29T02:45:31.683000", "content": "It is written (in \"Data\") that \"…scans from dozens of different CT machines have been reprocessed to use the run length encoded lossless compression format but retain other differences such as the number of bits per pixel, pixel range, and pixel representation\" Probably, we need to adapt to this.
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "2362356": "I have noticed that the training set is provided in the form of slices, whereas medical CT images are typically presented as 3D images.\n\nCT images can exhibit various characteristics depending on the scanning machine, such as voxel spacing. Voxel spacing refers to the distance between neighboring voxels in three spatial dimensions (x, y, and z axes) within CT images. It represents the actual size of a pixel in a CT image in physical space. After examining a few individual slices, I discovered that their voxel spacing differed, implying that resampling the CT image might be necessary. However, the dataset only provides slices and lacks continuity, which poses challenges for data processing.\n\nIn terms of the issues that inconsistent voxel spacing can cause in deep learning:\n1. Adjusting the spacing of a CT image can impact its overall pixel value and contextual information. Increasing spacing reduces the overall pixel value while enhancing contextual information, whereas decreasing spacing increases the overall pixel value while reducing contextual information. Therefore, striking the right balance between the amount of contextual information in the network's patch size and the preservation of image details is crucial for achieving optimal performance.\n\n2. Deep learning models typically perform convolution operations on localized regions of the input data and share weights to extract features. Inconsistencies in voxel spacing among input images can disrupt weight sharing, potentially affecting the model's performance.\n\n3. Introducing CT images with varying voxel spacing to the model may result in challenges during feature extraction and matching due to inconsistent correspondence between pixels. This, in turn, can impact the model's performance.\n\n\n\nMaybe I'm misunderstanding the task? It's not a 3d image mission, just a 2d?\n\n\n\n", "2364742": "It is common to see varying voxel dimensions and slice \"thickness/distance\" in CT. Machines are different and scanning protocols are different, depending on the anatomy/pathology in question. All CT modalities acquire a 3D 'volume' of data, then reconstruct the data into 2D slices in various \"planes\" .. or orientations, which are the images we can see. Those images can then be reconstructed into other planes or volumes depending on the clinical need.\n\nIdeally, we'd scan every patient at the highest resolution possible. But for CT, ionizing radiation makes that a bad idea. There's always a trade-off between radiation dosage and image detail.\n\nChanging voxel sizes, or pixel dimension/spacing data for resampling will distort the anatomy .. which is probably not a good idea.", "2363875": "It is written (in \"Data\") that \"...scans from dozens of different CT machines have been reprocessed to use the run length encoded lossless compression format but retain other differences such as the number of bits per pixel, pixel range, and pixel representation\" Probably, we need to adapt to this." } }