Electron cryogenic tomography (cryo-ET) is a microscopy technique that allows 3D visualization of cellular structures at near-atomic resolution. Cryo-ET has provided valuable insights into how proteins and macromolecular complexes function in their native environments, and several technology advances have automated and accelerated several components of the workflow. However, a significant bottleneck occurs in the manual annotation of each 3D image, or tomogram. In this challenge, you’ll be responsible for developing a machine learning algorithm to automate this annotation process for flagellar motors.
\nA variety of single-celled organisms use long, helical fibers called flagella to move from place to place. Each flagellum is quickly rotated by a cellular machine called the flagellar motor, and this rotation propels the cell forward. The flagellar motor spans across the two membranes of the cell, and it is made up of several distinct rings of proteins: Rotor rings, dubbed the P-ring and the L-ring, rotate the flagellum fiber. Stator rings, called the C-ring and MS-ring, are stationary elements of the motor that provide structural stability for the rotor. Flagellar motors are comparatively large structures, making them relatively simple to identify in tomograms. You can learn more about flagellar motors, and what a tomogram looks like, here: https://www.cellstructureatlas.org/6-2-flagellar-motor.html.
\nYour goal is to develop machine learning algorithms capable of identifying and locating the presence of flagellar motors in a set of 3D tomograms. The training set contains 817 tomograms of a variety of different species, each with 0, 1, or multiple motors. Your algorithm will need to differentiate between bacteria with and without motors. For bacteria with motors, your algorithm will need to be able to locate the coordinates of the motor on all three axes.
\nIn addition to competing for one of five cash prizes, your work will establish a benchmark for the field and establish the basis for future collaborations between the cryo-ET and ML communities. Eventually, we hope to develop an algorithm that will be able to detect a large variety of cellular machines, which will accelerate the cryo-ET pipeline and contribute to breakthroughs in the world’s understanding of cellular biology.
\nOur team will be active on the Discussion boards, so reach out with any questions you have, and we will help you out. We are excited to connect the cryo-ET and machine learning communities, and we are confident that this collaboration will lead to lasting impact on structural biology.
\nElectron cryogenic tomography (cryo-ET) is a microscopy technique that allows 3D visualization of cellular structures at near-atomic resolution. Cryo-ET has provided valuable insights into how proteins and macromolecular complexes function in their native environments, and several technology advances have automated and accelerated several components of the workflow. However, a significant bottleneck occurs in the manual annotation of each 3D image, or tomogram. In this challenge, you’ll be responsible for developing a machine learning algorithm to automate this annotation process for flagellar motors.
\nA variety of single-celled organisms use long, helical fibers called flagella to move from place to place. Each flagellum is quickly rotated by a cellular machine called the flagellar motor, and this rotation propels the cell forward. The flagellar motor spans across the two membranes of the cell, and it is made up of several distinct rings of proteins: Rotor rings, dubbed the P-ring and the L-ring, rotate the flagellum fiber. Stator rings, called the C-ring and MS-ring, are stationary elements of the motor that provide structural stability for the rotor. Flagellar motors are comparatively large structures, making them relatively simple to identify in tomograms. You can learn more about flagellar motors, and what a tomogram looks like, here: https://www.cellstructureatlas.org/6-2-flagellar-motor.html.
\nYour goal is to develop machine learning algorithms capable of identifying and locating the presence of flagellar motors in a set of 3D tomograms. The training set contains 817 tomograms of a variety of different species, each with 0, 1, or multiple motors. Your algorithm will need to differentiate between bacteria with and without motors. For bacteria with motors, your algorithm will need to be able to locate the coordinates of the motor on all three axes.
\nIn addition to competing for one of five cash prizes, your work will establish a benchmark for the field and establish the basis for future collaborations between the cryo-ET and ML communities. Eventually, we hope to develop an algorithm that will be able to detect a large variety of cellular machines, which will accelerate the cryo-ET pipeline and contribute to breakthroughs in the world’s understanding of cellular biology.
\nOur team will be active on the Discussion boards, so reach out with any questions you have, and we will help you out. We are excited to connect the cryo-ET and machine learning communities, and we are confident that this collaboration will lead to lasting impact on structural biology.
\nJust wanted to say kudos to the hosts (and all those responsible for the underlying materials – like the Atlas of Bacterial and Archaeal Structure), this competition has some top tier background knowledge resources, code resources, and tons of data. Thanks!
", "rawMarkdown": "Just wanted to say kudos to the hosts (and all those responsible for the underlying materials – like the Atlas of Bacterial and Archaeal Structure), this competition has some top tier background knowledge resources, code resources, and tons of data. Thanks!", "votes": 5 }, { "id": 3142603, "postDate": "2025-03-06T14:26:24.283Z", "content": "Thank you for hosting the competition! I just check the data and I have a question. Is Voxel Spacing provided for the test data? I can see those parameters provided in the train data but seems nowhere to find in the test data.
", "rawMarkdown": "Thank you for hosting the competition! I just check the data and I have a question. Is **Voxel Spacing** provided for the test data? I can see those parameters provided in the train data but seems nowhere to find in the test data. ", "votes": 3, "replies": [ { "id": 3142716, "postDate": "2025-03-06T16:17:43.150Z", "content": "Voxel spacing is not currently provided for the test data--best of luck on the competition!
", "rawMarkdown": "Voxel spacing is not currently provided for the test data--best of luck on the competition!", "votes": 5, "replies": [ { "id": 3142843, "postDate": "2025-03-06T17:51:10.190Z", "content": "What is considered the 'centerpoint of a flagella's motor'? Would it be the center of the rod at the point where the stator meets the C ring, the center of the rod where the flagella hook connects to the rod, or the center of the C ring going inwards from the membrane?
", "rawMarkdown": "What is considered the 'centerpoint of a flagella's motor'? Would it be the center of the rod at the point where the stator meets the C ring, the center of the rod where the flagella hook connects to the rod, or the center of the C ring going inwards from the membrane?", "votes": 5 }, { "id": 3142853, "postDate": "2025-03-06T17:59:42.317Z", "content": "Additionally, is there a limit to how many flagellas are on each cell? Also, the cells dont possess any mutations that affect how the flagella motor is formed, right?
", "rawMarkdown": "Additionally, is there a limit to how many flagellas are on each cell? Also, the cells dont possess any mutations that affect how the flagella motor is formed, right?" }, { "id": 3144629, "postDate": "2025-03-08T16:37:02.107Z", "rawMarkdown": "", "isDeleted": true }, { "id": 3144632, "postDate": "2025-03-08T16:37:42.190Z", "content": "This is a great question—unfortunately I don’t know the exact answer. The dataset was annotated partly by BioPhysics grad students (who are likely to know the exact answer) and partly by physics and math students trained off of Grant Jensen’s cell atlas. The small variations in the annotations is accounted for by the 10000 angstrom radius of the sphere that is accepted as a true positive in the competition’s metric. You may also be able to figure this out by displaying the labels on top of the correct slice on a few of the tomograms in order to see which part of the motor is being labeled the center. I do that in my Demo Notebook and surely some other notebooks visualize it as well. Sorry I couldn’t be more helpful!
", "rawMarkdown": "This is a great question—unfortunately I don’t know the exact answer. The dataset was annotated partly by BioPhysics grad students (who are likely to know the exact answer) and partly by physics and math students trained off of Grant Jensen’s [cell atlas](https://www.cellstructureatlas.org/). The small variations in the annotations is accounted for by the 10000 angstrom radius of the sphere that is accepted as a true positive in the competition’s metric. You may also be able to figure this out by displaying the labels on top of the correct slice on a few of the tomograms in order to see which part of the motor is being labeled the center. I do that in my [Demo Notebook](https://www.kaggle.com/code/jacksonpond/demo-notebook) and surely some other notebooks visualize it as well. Sorry I couldn’t be more helpful!", "votes": 3 }, { "id": 3144634, "postDate": "2025-03-08T16:40:33.853Z", "content": "The most motors a cell in the training data has is ten. The test data contains only cells with 0 or 1 motors. The competition dataset contains regularly formed motors.
", "rawMarkdown": "The most motors a cell in the training data has is ten. The test data contains only cells with 0 or 1 motors. The competition dataset contains regularly formed motors.", "votes": 2 }, { "id": 3160144, "postDate": "2025-03-26T12:30:42.030Z", "content": "May I ask how all the slices are processed into the range of 0~255? As far as I know, raw cryo-ET tomograms often range much smaller, so are the slices preprocessed after the tomograms normalization, or are they just normed slice by slice only?
", "rawMarkdown": "May I ask how all the slices are processed into the range of 0~255? As far as I know, raw cryo-ET tomograms often range much smaller, so are the slices preprocessed after the tomograms normalization, or are they just normed slice by slice only? ", "votes": 3 }, { "id": 3168512, "postDate": "2025-04-02T14:13:18.370Z", "content": "Whether or not voxel spacing is provided for test data, does it really matter? We don't have labels for test data, so how do we know that it is not provided for test data? And, what is the significance of it in the context of we just have to locate the center of the motor? Thanks!
", "rawMarkdown": "Whether or not voxel spacing is provided for test data, does it really matter? We don't have labels for test data, so how do we know that it is not provided for test data? And, what is the significance of it in the context of we just have to locate the center of the motor? Thanks!" }, { "id": 3180147, "postDate": "2025-04-16T07:59:18.433Z", "content": "The voxel spacing is VERY relevant because it allows to reshape the images in a way that objects have a constant size. This is standard procedure in 3D medical images and I believe here as well. I would very much like to have the voxel spacing at test time as this would make data preprocessing much easier!
", "rawMarkdown": "The voxel spacing is VERY relevant because it allows to reshape the images in a way that objects have a constant size. This is standard procedure in 3D medical images and I believe here as well. I would very much like to have the voxel spacing at test time as this would make data preprocessing much easier!", "votes": 1 }, { "id": 3180148, "postDate": "2025-04-16T08:01:27.667Z", "content": "I would like to support this request by @yksinyoung . When using external data like the one generously provided by @brendanartley it is essential to be consistent with the competition data. I am not sure how he picked his normalization strategy, could be an educated guess for all we know
", "rawMarkdown": "I would like to support this request by @yksinyoung . When using external data like the one generously provided by @brendanartley it is essential to be consistent with the competition data. I am not sure how he picked his normalization strategy, could be an educated guess for all we know" } ] } ] }, { "id": 3141933, "postDate": "2025-03-06T01:43:46.900Z", "content": "Thanks for hosting! I didn't see this mentioned anywhere, are the x and y coordinates with respect to the bottom left or top left of the image?
\nEdit: A quick look at the data it appears that Axis 1 is actually the y coordinate, Axis 2 is the x-coordinate and the origin is the top left of the image.
", "rawMarkdown": "Thanks for hosting! I didn't see this mentioned anywhere, are the x and y coordinates with respect to the bottom left or top left of the image?\n\nEdit: A quick look at the data it appears that Axis 1 is actually the y coordinate, Axis 2 is the x-coordinate and the origin is the top left of the image.", "votes": 4, "replies": [ { "id": 3142451, "postDate": "2025-03-06T11:57:34.080Z", "content": "Now, I am really confused.
", "rawMarkdown": "Now, I am really confused.", "replies": [ { "id": 3142486, "postDate": "2025-03-06T12:41:19.250Z", "content": "If you load the volume with shape (depth, height, width), then the axis-0, axis-1, axis-2 indices are correct. The description on the data page is just confusing because they call axis-1 \"x\" and axis-2 \"y\", when really it should be the other way around.
", "rawMarkdown": "If you load the volume with shape (depth, height, width), then the axis-0, axis-1, axis-2 indices are correct. The description on the data page is just confusing because they call axis-1 \"x\" and axis-2 \"y\", when really it should be the other way around." } ] }, { "id": 3142722, "postDate": "2025-03-06T16:21:11.630Z", "content": "Yes, thank you for pointing this out. It's helpful to me to think of the stack of images as an array of voxels, so the coordinates are in the order (depth, height, width)--or what we normally think of as (z, y, x). And they are with respect to the top left corner. We will correct/clarify this on the data page!
", "rawMarkdown": "Yes, thank you for pointing this out. It's helpful to me to think of the stack of images as an array of voxels, so the coordinates are in the order (depth, height, width)--or what we normally think of as (z, y, x). And they are with respect to the top left corner. We will correct/clarify this on the data page!", "votes": 6 } ] }, { "id": 3148913, "postDate": "2025-03-13T16:20:47.600Z", "content": "Thanks for such a clear explanation and direction!
", "rawMarkdown": "Thanks for such a clear explanation and direction!", "votes": 1 }, { "id": 3147342, "postDate": "2025-03-12T00:21:59.577Z", "content": "Quick Question About Z-Axis in the Dataset
\nHi Braxton,
\nHope you’re doing great! Thanks for putting together such an amazing dataset for the BYU Flagellar Motors challenge.
\nI’ve been digging into the tomograms and noticed something I’m a bit unsure about: the voxel spacing. The Voxel spacing column in the dataset gives a single value, but I can’t tell if that applies equally to all three axes (z, y, x) or if the z-axis is handled differently (like slices being spaced further apart or something).
\nThis matters because, in machine learning, keeping things grounded in real-world units (like Ångströms) helps models make better predictions. If the z-axis is scaled differently and I don’t account for that, it could throw off the model's ability to measure distances accurately, which would be a bummer for performance.
\nSo, could you clarify how the voxel spacing works for the z-axis compared to the x and y axes? And if there’s any extra info about how the tomograms map to real-world dimensions, that’d be awesome too!
\nThanks a ton for your help! And if you’ve got any insider tips for working with the dataset, I’m all ears.
", "rawMarkdown": "Quick Question About Z-Axis in the Dataset\n\nHi Braxton,\n\nHope you’re doing great! Thanks for putting together such an amazing dataset for the BYU Flagellar Motors challenge.\n\nI’ve been digging into the tomograms and noticed something I’m a bit unsure about: the voxel spacing. The Voxel spacing column in the dataset gives a single value, but I can’t tell if that applies equally to all three axes (z, y, x) or if the z-axis is handled differently (like slices being spaced further apart or something).\n\nThis matters because, in machine learning, keeping things grounded in real-world units (like Ångströms) helps models make better predictions. If the z-axis is scaled differently and I don’t account for that, it could throw off the model's ability to measure distances accurately, which would be a bummer for performance.\n\nSo, could you clarify how the voxel spacing works for the z-axis compared to the x and y axes? And if there’s any extra info about how the tomograms map to real-world dimensions, that’d be awesome too!\n\nThanks a ton for your help! And if you’ve got any insider tips for working with the dataset, I’m all ears.", "votes": 1, "replies": [ { "id": 3147344, "postDate": "2025-03-12T00:28:11.100Z", "content": "Just noticed @andreizamfir 's message: This is under the assumption that the voxel spacing is indeed isotropic. Thanks, Andrei!
", "rawMarkdown": "Just noticed @andreizamfir 's message: This is under the assumption that the voxel spacing is indeed isotropic. Thanks, Andrei!", "votes": 1 } ] }, { "id": 3143322, "postDate": "2025-03-07T05:24:12.873Z", "content": "I'm glad to host this competition. May I ask about the approximate amount of test data? This may facilitate us in controlling the running time.
", "rawMarkdown": "I'm glad to host this competition. May I ask about the approximate amount of test data? This may facilitate us in controlling the running time.", "votes": 1 }, { "id": 3145290, "postDate": "2025-03-09T16:52:14.147Z", "content": "What are the units of the axes? Are they directly expressed in Armstrongs, or should the conversion factor be used?
", "rawMarkdown": "What are the units of the axes? Are they directly expressed in Armstrongs, or should the conversion factor be used?", "votes": 2, "replies": [ { "id": 3145311, "postDate": "2025-03-09T17:30:48.293Z", "content": "The units of the axes are pixels, and for each tomography, you get a voxel spacing expressed in angstroms (as an intuition, think of how many angstroms is the spacing between 2 pixels on the same axis). If you multiply them, you get the tomography size in angstroms.
\nFor example, for tomo_003acc, with a voxel spacing of 6.5 and z, y, x = 500, 1912, 1847, the spatial dimensions in angstroms become 3250, 12428, 12005.5.
\nThis is, of course, under the assumption that the voxel spacing is indeed isotropic.
", "rawMarkdown": "The units of the axes are pixels, and for each tomography, you get a voxel spacing expressed in angstroms (as an intuition, think of how many angstroms is the spacing between 2 pixels on the same axis). If you multiply them, you get the tomography size in angstroms. \n\nFor example, for tomo_003acc, with a voxel spacing of 6.5 and z, y, x = 500, 1912, 1847, the spatial dimensions in angstroms become 3250, 12428, 12005.5. \n\nThis is, of course, under the assumption that the voxel spacing is indeed isotropic.", "votes": 5 } ] }, { "id": 3142960, "postDate": "2025-03-06T19:35:06.860Z", "content": "Did you draw the training and test sets randomly from the same initial dataset, or is there some population shift between the two?
\nIf there is a population shift between training and test, is there also one between public test and private test?
", "rawMarkdown": "Did you draw the training and test sets randomly from the same initial dataset, or is there some population shift between the two?\n\nIf there is a population shift between training and test, is there also one between public test and private test?", "votes": 2, "replies": [ { "id": 3142976, "postDate": "2025-03-06T19:50:03.603Z", "content": "I know there is a population shift between the train and test set as to avoid metadata getting in the way of the competition integrity.
", "rawMarkdown": "I know there is a population shift between the train and test set as to avoid metadata getting in the way of the competition integrity. ", "votes": 4, "replies": [ { "id": 3144388, "postDate": "2025-03-08T10:00:14.147Z", "content": "Do you mean training and test set images are acquired using different machines and configurations?
", "rawMarkdown": "Do you mean training and test set images are acquired using different machines and configurations?", "votes": 3 }, { "id": 3145721, "postDate": "2025-03-10T07:45:17.353Z", "content": "Yes I do. They were also sampled by different people.
", "rawMarkdown": "Yes I do. They were also sampled by different people.", "votes": 8 } ] } ] }, { "id": 3191926, "postDate": "2025-05-02T09:30:49.163Z", "content": "@andrewjdarley @jacksonpond Not sure if I missed this discussion anywhere, but are the images in BYU - Locating Bacterial Flagellar Motors (V2) a subset of this competition data? If not are we allowed to use this dataset for training models for this competition?
", "rawMarkdown": "@andrewjdarley @jacksonpond Not sure if I missed this discussion anywhere, but are the images in [BYU - Locating Bacterial Flagellar Motors (V2)](https://www.kaggle.com/competitions/byu-locating-bacterial-flagellar-motors-v2/data) a subset of this competition data? If not are we allowed to use this dataset for training models for this competition?", "replies": [ { "id": 3197797, "postDate": "2025-05-08T15:47:04.990Z", "content": "The images in that competition are a subset of this competition data that's correct. The data found there is unlikely to be helpful as it was more of a test run of this competition meant to get the team familiar with the ropes. It is a much simpler dataset containing only a single 2D slice for each instance
", "rawMarkdown": "The images in that competition are a subset of this competition data that's correct. The data found there is unlikely to be helpful as it was more of a test run of this competition meant to get the team familiar with the ropes. It is a much simpler dataset containing only a single 2D slice for each instance" } ] }, { "id": 3185201, "postDate": "2025-04-23T02:09:37.233Z", "content": "Thank you for the explanation. It takes about 7 hours to run my notebook—does this cover all of the test data or just 30% of it? Is training the model part of that 12 hours?
", "rawMarkdown": "Thank you for the explanation. It takes about 7 hours to run my notebook—does this cover all of the test data or just 30% of it? Is training the model part of that 12 hours?", "replies": [ { "id": 3187268, "postDate": "2025-04-25T18:19:40.847Z", "content": "Every submission will run on all the test data, not just the public data. If you train your model in another notebook or on your own machine and upload the weights as input into your submission notebook you don't have to use the allotted hours for training.
", "rawMarkdown": "Every submission will run on all the test data, not just the public data. If you train your model in another notebook or on your own machine and upload the weights as input into your submission notebook you don't have to use the allotted hours for training. " } ] }, { "id": 3173366, "postDate": "2025-04-07T19:54:42.380Z", "content": "Good day! Thank you for hosting this competition!
\nCould you please tell me how many tomograms there are in the public and private test sets? I would like to estimate the runtime of my notebook to make sure it fits within the time limit.
\nAlso, could you please clarify if I understand correctly: is it true that there can be either zero or one motor in each tomogram in both the public and private test sets?
\nThank you! Looking forward to your reply.
", "rawMarkdown": "Good day! Thank you for hosting this competition! \n\nCould you please tell me how many tomograms there are in the public and private test sets? I would like to estimate the runtime of my notebook to make sure it fits within the time limit.\n\nAlso, could you please clarify if I understand correctly: is it true that there can be either zero or one motor in each tomogram in **both** the public and private test sets?\n\nThank you! Looking forward to your reply.", "replies": [ { "id": 3174077, "postDate": "2025-04-08T16:48:05.190Z", "content": "Hey, thanks for your comment, Yaroslav!
\n\n\nCould you please tell me how many tomograms there are in the public and private test sets? I would like to estimate the runtime of my notebook to make sure it fits within the time limit.
\n
There are 648 train tomograms, and approximately 900 test tomograms (see the data page).
\n\n\nAlso, could you please clarify if I understand correctly: is it true that there can be either zero or one motor in each tomogram in both the public and private test sets?
\n
Yes this is correct
", "rawMarkdown": "Hey, thanks for your comment, Yaroslav!\n\n>Could you please tell me how many tomograms there are in the public and private test sets? I would like to estimate the runtime of my notebook to make sure it fits within the time limit.\n\nThere are 648 train tomograms, and approximately 900 test tomograms (see the data page).\n\n>Also, could you please clarify if I understand correctly: is it true that there can be either zero or one motor in each tomogram in both the public and private test sets?\n\nYes this is correct", "votes": 1, "replies": [ { "id": 3185221, "postDate": "2025-04-23T02:49:59.193Z", "content": "But there is more than one motor per tomogram in the training set.
", "rawMarkdown": "But there is more than one motor per tomogram in the training set. " } ] } ] }, { "id": 3156189, "postDate": "2025-03-21T20:53:10.463Z", "content": "Hi!. I want to applause for the great competition you provided. I have several question If you don't mind.
\nHow is the data collected? Did you do some denoising? Also, why use F beta 2 for the metrics, is it related to how the data are collected?
\nThank you.
", "rawMarkdown": "Hi!. I want to applause for the great competition you provided. I have several question If you don't mind.\n\nHow is the data collected? Did you do some denoising? Also, why use F beta 2 for the metrics, is it related to how the data are collected?\n\nThank you. " }, { "id": 3143531, "postDate": "2025-03-07T10:06:18.693Z", "content": "hola¡¡¡ no se si mi codigo es eficiente… me gustara sasberlo… gracias.
", "rawMarkdown": "hola¡¡¡ no se si mi codigo es eficiente... me gustara sasberlo... gracias." } ], "comments": [ { "id": 3146193, "author_name": "Darien Schettler", "author_url": "", "post_date": "2025-03-10T16:14:25.983000", "content": "Just wanted to say kudos to the hosts (and all those responsible for the underlying materials – like the Atlas of Bacterial and Archaeal Structure), this competition has some top tier background knowledge resources, code resources, and tons of data. Thanks!
", "votes": 5, "replies": [] }, { "id": 3142603, "author_name": "Yksin Young", "author_url": "", "post_date": "2025-03-06T14:26:24.283000", "content": "Thank you for hosting the competition! I just check the data and I have a question. Is Voxel Spacing provided for the test data? I can see those parameters provided in the train data but seems nowhere to find in the test data.
", "votes": 3, "replies": [ { "id": 3142716, "author_name": "Jackson Pond", "author_url": "", "post_date": "2025-03-06T16:17:43.150000", "content": "Voxel spacing is not currently provided for the test data--best of luck on the competition!
", "votes": 5, "replies": [ { "id": 3142843, "author_name": "Quin O’Malley", "author_url": "", "post_date": "2025-03-06T17:51:10.190000", "content": "What is considered the 'centerpoint of a flagella's motor'? Would it be the center of the rod at the point where the stator meets the C ring, the center of the rod where the flagella hook connects to the rod, or the center of the C ring going inwards from the membrane?
", "votes": 5, "replies": [] }, { "id": 3142853, "author_name": "Quin O’Malley", "author_url": "", "post_date": "2025-03-06T17:59:42.317000", "content": "Additionally, is there a limit to how many flagellas are on each cell? Also, the cells dont possess any mutations that affect how the flagella motor is formed, right?
", "votes": 0, "replies": [] }, { "id": 3144629, "author_name": "", "author_url": "", "post_date": "2025-03-08T16:37:02.107000", "content": "", "votes": 0, "replies": [] }, { "id": 3144632, "author_name": "Jackson Pond", "author_url": "", "post_date": "2025-03-08T16:37:42.190000", "content": "This is a great question—unfortunately I don’t know the exact answer. The dataset was annotated partly by BioPhysics grad students (who are likely to know the exact answer) and partly by physics and math students trained off of Grant Jensen’s cell atlas. The small variations in the annotations is accounted for by the 10000 angstrom radius of the sphere that is accepted as a true positive in the competition’s metric. You may also be able to figure this out by displaying the labels on top of the correct slice on a few of the tomograms in order to see which part of the motor is being labeled the center. I do that in my Demo Notebook and surely some other notebooks visualize it as well. Sorry I couldn’t be more helpful!
", "votes": 3, "replies": [] }, { "id": 3144634, "author_name": "Jackson Pond", "author_url": "", "post_date": "2025-03-08T16:40:33.853000", "content": "The most motors a cell in the training data has is ten. The test data contains only cells with 0 or 1 motors. The competition dataset contains regularly formed motors.
", "votes": 2, "replies": [] }, { "id": 3160144, "author_name": "Yksin Young", "author_url": "", "post_date": "2025-03-26T12:30:42.030000", "content": "May I ask how all the slices are processed into the range of 0~255? As far as I know, raw cryo-ET tomograms often range much smaller, so are the slices preprocessed after the tomograms normalization, or are they just normed slice by slice only?
", "votes": 3, "replies": [] }, { "id": 3168512, "author_name": "Meshkat Ahammed", "author_url": "", "post_date": "2025-04-02T14:13:18.370000", "content": "Whether or not voxel spacing is provided for test data, does it really matter? We don't have labels for test data, so how do we know that it is not provided for test data? And, what is the significance of it in the context of we just have to locate the center of the motor? Thanks!
", "votes": 0, "replies": [] }, { "id": 3180147, "author_name": "FabianIsensee", "author_url": "", "post_date": "2025-04-16T07:59:18.433000", "content": "The voxel spacing is VERY relevant because it allows to reshape the images in a way that objects have a constant size. This is standard procedure in 3D medical images and I believe here as well. I would very much like to have the voxel spacing at test time as this would make data preprocessing much easier!
", "votes": 1, "replies": [] }, { "id": 3180148, "author_name": "FabianIsensee", "author_url": "", "post_date": "2025-04-16T08:01:27.667000", "content": "I would like to support this request by @yksinyoung . When using external data like the one generously provided by @brendanartley it is essential to be consistent with the competition data. I am not sure how he picked his normalization strategy, could be an educated guess for all we know
", "votes": 0, "replies": [] } ] } ] }, { "id": 3141933, "author_name": "John Payne", "author_url": "", "post_date": "2025-03-06T01:43:46.900000", "content": "Thanks for hosting! I didn't see this mentioned anywhere, are the x and y coordinates with respect to the bottom left or top left of the image?
\nEdit: A quick look at the data it appears that Axis 1 is actually the y coordinate, Axis 2 is the x-coordinate and the origin is the top left of the image.
", "votes": 4, "replies": [ { "id": 3142451, "author_name": "Sai Ram_3", "author_url": "", "post_date": "2025-03-06T11:57:34.080000", "content": "Now, I am really confused.
", "votes": 0, "replies": [ { "id": 3142486, "author_name": "John Payne", "author_url": "", "post_date": "2025-03-06T12:41:19.250000", "content": "If you load the volume with shape (depth, height, width), then the axis-0, axis-1, axis-2 indices are correct. The description on the data page is just confusing because they call axis-1 \"x\" and axis-2 \"y\", when really it should be the other way around.
", "votes": 0, "replies": [] } ] }, { "id": 3142722, "author_name": "Jackson Pond", "author_url": "", "post_date": "2025-03-06T16:21:11.630000", "content": "Yes, thank you for pointing this out. It's helpful to me to think of the stack of images as an array of voxels, so the coordinates are in the order (depth, height, width)--or what we normally think of as (z, y, x). And they are with respect to the top left corner. We will correct/clarify this on the data page!
", "votes": 6, "replies": [] } ] }, { "id": 3148913, "author_name": "Carlo Finnegan", "author_url": "", "post_date": "2025-03-13T16:20:47.600000", "content": "Thanks for such a clear explanation and direction!
", "votes": 1, "replies": [] }, { "id": 3147342, "author_name": "Jeff.271", "author_url": "", "post_date": "2025-03-12T00:21:59.577000", "content": "Quick Question About Z-Axis in the Dataset
\nHi Braxton,
\nHope you’re doing great! Thanks for putting together such an amazing dataset for the BYU Flagellar Motors challenge.
\nI’ve been digging into the tomograms and noticed something I’m a bit unsure about: the voxel spacing. The Voxel spacing column in the dataset gives a single value, but I can’t tell if that applies equally to all three axes (z, y, x) or if the z-axis is handled differently (like slices being spaced further apart or something).
\nThis matters because, in machine learning, keeping things grounded in real-world units (like Ångströms) helps models make better predictions. If the z-axis is scaled differently and I don’t account for that, it could throw off the model's ability to measure distances accurately, which would be a bummer for performance.
\nSo, could you clarify how the voxel spacing works for the z-axis compared to the x and y axes? And if there’s any extra info about how the tomograms map to real-world dimensions, that’d be awesome too!
\nThanks a ton for your help! And if you’ve got any insider tips for working with the dataset, I’m all ears.
", "votes": 1, "replies": [ { "id": 3147344, "author_name": "Jeff.271", "author_url": "", "post_date": "2025-03-12T00:28:11.100000", "content": "Just noticed @andreizamfir 's message: This is under the assumption that the voxel spacing is indeed isotropic. Thanks, Andrei!
", "votes": 1, "replies": [] } ] }, { "id": 3143322, "author_name": "wpl", "author_url": "", "post_date": "2025-03-07T05:24:12.873000", "content": "I'm glad to host this competition. May I ask about the approximate amount of test data? This may facilitate us in controlling the running time.
", "votes": 1, "replies": [] }, { "id": 3145290, "author_name": "elmafia", "author_url": "", "post_date": "2025-03-09T16:52:14.147000", "content": "What are the units of the axes? Are they directly expressed in Armstrongs, or should the conversion factor be used?
", "votes": 2, "replies": [ { "id": 3145311, "author_name": "Andrei Zamfir", "author_url": "", "post_date": "2025-03-09T17:30:48.293000", "content": "The units of the axes are pixels, and for each tomography, you get a voxel spacing expressed in angstroms (as an intuition, think of how many angstroms is the spacing between 2 pixels on the same axis). If you multiply them, you get the tomography size in angstroms.
\nFor example, for tomo_003acc, with a voxel spacing of 6.5 and z, y, x = 500, 1912, 1847, the spatial dimensions in angstroms become 3250, 12428, 12005.5.
\nThis is, of course, under the assumption that the voxel spacing is indeed isotropic.
", "votes": 5, "replies": [] } ] }, { "id": 3142960, "author_name": "Jeroen Cottaar", "author_url": "", "post_date": "2025-03-06T19:35:06.860000", "content": "Did you draw the training and test sets randomly from the same initial dataset, or is there some population shift between the two?
\nIf there is a population shift between training and test, is there also one between public test and private test?
", "votes": 2, "replies": [ { "id": 3142976, "author_name": "Andrew Darley", "author_url": "", "post_date": "2025-03-06T19:50:03.603000", "content": "I know there is a population shift between the train and test set as to avoid metadata getting in the way of the competition integrity.
", "votes": 4, "replies": [ { "id": 3144388, "author_name": "Gunes Evitan", "author_url": "", "post_date": "2025-03-08T10:00:14.147000", "content": "Do you mean training and test set images are acquired using different machines and configurations?
", "votes": 3, "replies": [] }, { "id": 3145721, "author_name": "Andrew Darley", "author_url": "", "post_date": "2025-03-10T07:45:17.353000", "content": "Yes I do. They were also sampled by different people.
", "votes": 8, "replies": [] } ] } ] }, { "id": 3191926, "author_name": "Yovin Yahathugoda", "author_url": "", "post_date": "2025-05-02T09:30:49.163000", "content": "@andrewjdarley @jacksonpond Not sure if I missed this discussion anywhere, but are the images in BYU - Locating Bacterial Flagellar Motors (V2) a subset of this competition data? If not are we allowed to use this dataset for training models for this competition?
", "votes": 0, "replies": [ { "id": 3197797, "author_name": "Andrew Darley", "author_url": "", "post_date": "2025-05-08T15:47:04.990000", "content": "The images in that competition are a subset of this competition data that's correct. The data found there is unlikely to be helpful as it was more of a test run of this competition meant to get the team familiar with the ropes. It is a much simpler dataset containing only a single 2D slice for each instance
", "votes": 0, "replies": [] } ] }, { "id": 3185201, "author_name": "Hossein", "author_url": "", "post_date": "2025-04-23T02:09:37.233000", "content": "Thank you for the explanation. It takes about 7 hours to run my notebook—does this cover all of the test data or just 30% of it? Is training the model part of that 12 hours?
", "votes": 0, "replies": [ { "id": 3187268, "author_name": "Andrew Darley", "author_url": "", "post_date": "2025-04-25T18:19:40.847000", "content": "Every submission will run on all the test data, not just the public data. If you train your model in another notebook or on your own machine and upload the weights as input into your submission notebook you don't have to use the allotted hours for training.
", "votes": 0, "replies": [] } ] }, { "id": 3173366, "author_name": "Yaroslav Nikolaev", "author_url": "", "post_date": "2025-04-07T19:54:42.380000", "content": "Good day! Thank you for hosting this competition!
\nCould you please tell me how many tomograms there are in the public and private test sets? I would like to estimate the runtime of my notebook to make sure it fits within the time limit.
\nAlso, could you please clarify if I understand correctly: is it true that there can be either zero or one motor in each tomogram in both the public and private test sets?
\nThank you! Looking forward to your reply.
", "votes": 0, "replies": [ { "id": 3174077, "author_name": "Andrew Darley", "author_url": "", "post_date": "2025-04-08T16:48:05.190000", "content": "Hey, thanks for your comment, Yaroslav!
\n\n\nCould you please tell me how many tomograms there are in the public and private test sets? I would like to estimate the runtime of my notebook to make sure it fits within the time limit.
\n
There are 648 train tomograms, and approximately 900 test tomograms (see the data page).
\n\n\nAlso, could you please clarify if I understand correctly: is it true that there can be either zero or one motor in each tomogram in both the public and private test sets?
\n
Yes this is correct
", "votes": 1, "replies": [ { "id": 3185221, "author_name": "Meshkat Ahammed", "author_url": "", "post_date": "2025-04-23T02:49:59.193000", "content": "But there is more than one motor per tomogram in the training set.
", "votes": 0, "replies": [] } ] } ] }, { "id": 3156189, "author_name": "bagas.jwnt", "author_url": "", "post_date": "2025-03-21T20:53:10.463000", "content": "Hi!. I want to applause for the great competition you provided. I have several question If you don't mind.
\nHow is the data collected? Did you do some denoising? Also, why use F beta 2 for the metrics, is it related to how the data are collected?
\nThank you.
", "votes": 0, "replies": [] }, { "id": 3143531, "author_name": "Fuego Y Agua", "author_url": "", "post_date": "2025-03-07T10:06:18.693000", "content": "hola¡¡¡ no se si mi codigo es eficiente… me gustara sasberlo… gracias.
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "3139832": "##We’re excited to welcome you to this competition, which has potential to make a significant impact on our understanding of structural biology.\n\n##About the competition\nElectron cryogenic tomography (cryo-ET) is a microscopy technique that allows 3D visualization of cellular structures at near-atomic resolution. Cryo-ET has provided valuable insights into how proteins and macromolecular complexes function in their native environments, and several technology advances have automated and accelerated several components of the workflow. However, a significant bottleneck occurs in the manual annotation of each 3D image, or tomogram. In this challenge, you’ll be responsible for developing a machine learning algorithm to automate this annotation process for flagellar motors.\n\n##What’s a flagellar motor? \nA variety of single-celled organisms use long, helical fibers called flagella to move from place to place. Each flagellum is quickly rotated by a cellular machine called the flagellar motor, and this rotation propels the cell forward. The flagellar motor spans across the two membranes of the cell, and it is made up of several distinct rings of proteins: Rotor rings, dubbed the P-ring and the L-ring, rotate the flagellum fiber. Stator rings, called the C-ring and MS-ring, are stationary elements of the motor that provide structural stability for the rotor. Flagellar motors are comparatively large structures, making them relatively simple to identify in tomograms. You can learn more about flagellar motors, and what a tomogram looks like, here: https://www.cellstructureatlas.org/6-2-flagellar-motor.html. \n\n##About the challenge\nYour goal is to develop machine learning algorithms capable of identifying and locating the presence of flagellar motors in a set of 3D tomograms. The training set contains 817 tomograms of a variety of different species, each with 0, 1, or multiple motors. Your algorithm will need to differentiate between bacteria with and without motors. For bacteria with motors, your algorithm will need to be able to locate the coordinates of the motor on all three axes. \n\n##Why participate? \nIn addition to competing for one of five cash prizes, your work will establish a benchmark for the field and establish the basis for future collaborations between the cryo-ET and ML communities. Eventually, we hope to develop an algorithm that will be able to detect a large variety of cellular machines, which will accelerate the cryo-ET pipeline and contribute to breakthroughs in the world’s understanding of cellular biology. \n\n##Talk with us\nOur team will be active on the Discussion boards, so reach out with any questions you have, and we will help you out. We are excited to connect the cryo-ET and machine learning communities, and we are confident that this collaboration will lead to lasting impact on structural biology. \n\n##Learn more\n- The [Atlas of Bacterial and Archaeal Cell Structure](https://www.cellstructureatlas.org/index.html) contains the best of the Jensen lab’s tomograms, annotated with features of interest and accompanied by explanatory text.\n- [Starting Notebooks](https://www.kaggle.com/code/andrewjdarley/byu-biophysics-group-sample-solution)\n- [CZI Cryo-ET Data Portal](https://cryoetdataportal.czscience.com/)", "3146193": "Just wanted to say kudos to the hosts (and all those responsible for the underlying materials – like the Atlas of Bacterial and Archaeal Structure), this competition has some top tier background knowledge resources, code resources, and tons of data. Thanks!", "3142603": "Thank you for hosting the competition! I just check the data and I have a question. Is **Voxel Spacing** provided for the test data? I can see those parameters provided in the train data but seems nowhere to find in the test data. ", "3141933": "Thanks for hosting! I didn't see this mentioned anywhere, are the x and y coordinates with respect to the bottom left or top left of the image?\n\nEdit: A quick look at the data it appears that Axis 1 is actually the y coordinate, Axis 2 is the x-coordinate and the origin is the top left of the image.", "3148913": "Thanks for such a clear explanation and direction!", "3147342": "Quick Question About Z-Axis in the Dataset\n\nHi Braxton,\n\nHope you’re doing great! Thanks for putting together such an amazing dataset for the BYU Flagellar Motors challenge.\n\nI’ve been digging into the tomograms and noticed something I’m a bit unsure about: the voxel spacing. The Voxel spacing column in the dataset gives a single value, but I can’t tell if that applies equally to all three axes (z, y, x) or if the z-axis is handled differently (like slices being spaced further apart or something).\n\nThis matters because, in machine learning, keeping things grounded in real-world units (like Ångströms) helps models make better predictions. If the z-axis is scaled differently and I don’t account for that, it could throw off the model's ability to measure distances accurately, which would be a bummer for performance.\n\nSo, could you clarify how the voxel spacing works for the z-axis compared to the x and y axes? And if there’s any extra info about how the tomograms map to real-world dimensions, that’d be awesome too!\n\nThanks a ton for your help! And if you’ve got any insider tips for working with the dataset, I’m all ears.", "3143322": "I'm glad to host this competition. May I ask about the approximate amount of test data? This may facilitate us in controlling the running time.", "3145290": "What are the units of the axes? Are they directly expressed in Armstrongs, or should the conversion factor be used?", "3142960": "Did you draw the training and test sets randomly from the same initial dataset, or is there some population shift between the two?\n\nIf there is a population shift between training and test, is there also one between public test and private test?", "3191926": "@andrewjdarley @jacksonpond Not sure if I missed this discussion anywhere, but are the images in [BYU - Locating Bacterial Flagellar Motors (V2)](https://www.kaggle.com/competitions/byu-locating-bacterial-flagellar-motors-v2/data) a subset of this competition data? If not are we allowed to use this dataset for training models for this competition?", "3185201": "Thank you for the explanation. It takes about 7 hours to run my notebook—does this cover all of the test data or just 30% of it? Is training the model part of that 12 hours?", "3173366": "Good day! Thank you for hosting this competition! \n\nCould you please tell me how many tomograms there are in the public and private test sets? I would like to estimate the runtime of my notebook to make sure it fits within the time limit.\n\nAlso, could you please clarify if I understand correctly: is it true that there can be either zero or one motor in each tomogram in **both** the public and private test sets?\n\nThank you! Looking forward to your reply.", "3156189": "Hi!. I want to applause for the great competition you provided. I have several question If you don't mind.\n\nHow is the data collected? Did you do some denoising? Also, why use F beta 2 for the metrics, is it related to how the data are collected?\n\nThank you. ", "3143531": "hola¡¡¡ no se si mi codigo es eficiente... me gustara sasberlo... gracias." } }