Hello, CZ imaging team.
\nI read your introduction thesis of this competition, and find your part of Establishing “ground truth” labels
\n\n\nA more refined and expansive set of picks was generated using two deep learning approaches. One of these, DeepFindET (an adaptation of DeepFinder12), uses a Residual U-Net architecture with three 3D convolutional layers and a receptive field size of 680 Å to predict segmentation masks from annotated tomograms. The other, CellCanvas, involves a multi-step pipeline to segment tomograms. In the first step, a Swin UNETR model pre-trained on computed tomography data42 generates a voxel-wise embedding for each tomogram. In the second step, an interactively trained XGBoost classifier transforms clusters in embedding space into multi-class segmentation masks so that each voxel of the tomogram is assigned a probability for each class label. For initial training data, synthetic tomograms were generated in PolNet43 with a thickness of 180 nm and the five target particles, Beta-amylase, and membranes randomly distributed throughout the sample volume.
\n
Would you mind giving more details of 'an interactively trained XGBoost classifier transforms clusters in embedding space into multi-class segmentation masks so that each voxel of the tomogram is assigned a probability for each class label.' I really want to know how to use XGBoost classifier to make it.
\nMany thanks.
", "messages": [ { "id": "3056777", "postDate": "11/27/2024 11:13:39", "content": "Hello, CZ imaging team.
\nI read your introduction thesis of this competition, and find your part of Establishing “ground truth” labels
\n\n\nA more refined and expansive set of picks was generated using two deep learning approaches. One of these, DeepFindET (an adaptation of DeepFinder12), uses a Residual U-Net architecture with three 3D convolutional layers and a receptive field size of 680 Å to predict segmentation masks from annotated tomograms. The other, CellCanvas, involves a multi-step pipeline to segment tomograms. In the first step, a Swin UNETR model pre-trained on computed tomography data42 generates a voxel-wise embedding for each tomogram. In the second step, an interactively trained XGBoost classifier transforms clusters in embedding space into multi-class segmentation masks so that each voxel of the tomogram is assigned a probability for each class label. For initial training data, synthetic tomograms were generated in PolNet43 with a thickness of 180 nm and the five target particles, Beta-amylase, and membranes randomly distributed throughout the sample volume.
\n
Would you mind giving more details of 'an interactively trained XGBoost classifier transforms clusters in embedding space into multi-class segmentation masks so that each voxel of the tomogram is assigned a probability for each class label.' I really want to know how to use XGBoost classifier to make it.
\nMany thanks.
", "rawMarkdown": "Hello, CZ imaging team. \n\nI read your [introduction thesis of this competition](https://www.biorxiv.org/content/10.1101/2024.11.04.621686v2), and find your part of Establishing “ground truth” labels\n\n>A more refined and expansive set of picks was generated using two deep learning approaches. One of these, DeepFindET (an adaptation of DeepFinder12), uses a Residual U-Net architecture with three 3D convolutional layers and a receptive field size of 680 Å to predict segmentation masks from annotated tomograms. The other, CellCanvas, involves a multi-step pipeline to segment tomograms. In the first step, a Swin UNETR model pre-trained on computed tomography data42 generates a voxel-wise embedding for each tomogram. In the second step, an interactively trained XGBoost classifier transforms clusters in embedding space into multi-class segmentation masks so that each voxel of the tomogram is assigned a probability for each class label. For initial training data, synthetic tomograms were generated in PolNet43 with a thickness of 180 nm and the five target particles, Beta-amylase, and membranes randomly distributed throughout the sample volume. \n\nWould you mind giving more details of 'an interactively trained XGBoost classifier transforms clusters in embedding space into multi-class segmentation masks so that each voxel of the tomogram is assigned a probability for each class label.' I really want to know how to use XGBoost classifier to make it.\n\nMany thanks.", "votes": null }, { "id": "3056811", "postDate": "11/27/2024 12:20:02", "content": "i do not know exactly how it is done in the paper, but here is some possibility:
\nextract feature from NN feature map:
\ntexture feature1 = NN feature map[xyz1]
\ntexture feature2 = NN feature map[xyz2]
you can compute some features like area, eccentricity, etc …
\nshape feature 1 = …
you can crop the original volume at coords
\ncrop = crop at xyz1
\ncompare texture or pixel feature at crop
all these features can be used as input for xgboost.
\nas an example:
\nClassification and Segmentation of Pulmonary Lesions in CT Images Using a Combined VGG-XGBoost Method, and an Integrated Fuzzy Clustering-Level Set Technique
\nhttps://arxiv.org/pdf/2101.00948
\nAfter reading the CT-Scan image, its features are automatically extracted by the VGG convolutional neural network.
\nThen, based on extracted features, it is classified using the XGBoost classifier. I…
Exactly. I am just curious of what features they extract for xgboost.
", "rawMarkdown": "Exactly. I am just curious of what features they extract for xgboost.", "votes": null }, { "id": "3056849", "postDate": "11/27/2024 13:11:54", "content": "Wavelet, fft , random pixel difference, cnn features, traditional texture statistics features
", "rawMarkdown": "Wavelet, fft , random pixel difference, cnn features, traditional texture statistics features", "votes": null } ], "comments": [ { "id": 3056811, "author_name": "hengck23", "author_url": "", "post_date": "11/27/2024 12:20:02", "content": "i do not know exactly how it is done in the paper, but here is some possibility:
\nextract feature from NN feature map:
\ntexture feature1 = NN feature map[xyz1]
\ntexture feature2 = NN feature map[xyz2]
you can compute some features like area, eccentricity, etc …
\nshape feature 1 = …
you can crop the original volume at coords
\ncrop = crop at xyz1
\ncompare texture or pixel feature at crop
all these features can be used as input for xgboost.
\nas an example:
\nClassification and Segmentation of Pulmonary Lesions in CT Images Using a Combined VGG-XGBoost Method, and an Integrated Fuzzy Clustering-Level Set Technique
\nhttps://arxiv.org/pdf/2101.00948
\nAfter reading the CT-Scan image, its features are automatically extracted by the VGG convolutional neural network.
\nThen, based on extracted features, it is classified using the XGBoost classifier. I…
Exactly. I am just curious of what features they extract for xgboost.
", "votes": null, "replies": [ { "id": 3056849, "author_name": "hengck23", "author_url": "", "post_date": "11/27/2024 13:11:54", "content": "Wavelet, fft , random pixel difference, cnn features, traditional texture statistics features
", "votes": null, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "3056777": "Hello, CZ imaging team. \n\nI read your [introduction thesis of this competition](https://www.biorxiv.org/content/10.1101/2024.11.04.621686v2), and find your part of Establishing “ground truth” labels\n\n>A more refined and expansive set of picks was generated using two deep learning approaches. One of these, DeepFindET (an adaptation of DeepFinder12), uses a Residual U-Net architecture with three 3D convolutional layers and a receptive field size of 680 Å to predict segmentation masks from annotated tomograms. The other, CellCanvas, involves a multi-step pipeline to segment tomograms. In the first step, a Swin UNETR model pre-trained on computed tomography data42 generates a voxel-wise embedding for each tomogram. In the second step, an interactively trained XGBoost classifier transforms clusters in embedding space into multi-class segmentation masks so that each voxel of the tomogram is assigned a probability for each class label. For initial training data, synthetic tomograms were generated in PolNet43 with a thickness of 180 nm and the five target particles, Beta-amylase, and membranes randomly distributed throughout the sample volume. \n\nWould you mind giving more details of 'an interactively trained XGBoost classifier transforms clusters in embedding space into multi-class segmentation masks so that each voxel of the tomogram is assigned a probability for each class label.' I really want to know how to use XGBoost classifier to make it.\n\nMany thanks.", "3056811": "i do not know exactly how it is done in the paper, but here is some possibility:\n\n1. say you have already detected particles at coords =[ xyz1, xyz2 ...]\n- extract feature from NN feature map:\ntexture feature1 = NN feature map[xyz1]\ntexture feature2 = NN feature map[xyz2]\n\n- you can compute some features like area, eccentricity, etc ...\nshape feature 1 = ...\n\n- you can crop the original volume at coords\ncrop = crop at xyz1\ncompare texture or pixel feature at crop \n\n\nall these features can be used as input for xgboost.\n\nas an example:\nClassification and Segmentation of Pulmonary Lesions in CT Images Using a Combined VGG-XGBoost Method, and an Integrated Fuzzy Clustering-Level Set Technique\nhttps://arxiv.org/pdf/2101.00948\nAfter reading the CT-Scan image, its features are automatically extracted by the VGG convolutional neural network.\nThen, based on extracted features, it is classified using the XGBoost classifier. I...", "3056848": "Exactly. I am just curious of what features they extract for xgboost.", "3056849": "Wavelet, fft , random pixel difference, cnn features, traditional texture statistics features" }, "source": "meta" }