![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F650%2Fd52280577ad0c8424680e27623c74c29%2FScreenshot%202025-04-02%20162640.png?generation=1743607672151998&alt=media\")
Thansk for providing the starter notebook.
\nhttps://www.kaggle.com/code/elucidata2025/kaggle-starter-notebook
I found some spots (those red dots) being outside the purple/violet region when I zoom in. Is it normal?
\n(Wanted to post a comment on that notebook in the Code section, but found that it has been disabled.)
", "messages": [ { "id": 3154641, "postDate": "2025-03-20T07:30:56.040Z", "content": "Thansk for providing the starter notebook.
\nhttps://www.kaggle.com/code/elucidata2025/kaggle-starter-notebook
I found some spots (those red dots) being outside the purple/violet region when I zoom in. Is it normal?
\n(Wanted to post a comment on that notebook in the Code section, but found that it has been disabled.)
", "rawMarkdown": "Thansk for providing the starter notebook. \nhttps://www.kaggle.com/code/elucidata2025/kaggle-starter-notebook\n\nI found some spots (those red dots) being outside the purple/violet region when I zoom in. Is it normal?\n\n(Wanted to post a comment on that notebook in the Code section, but found that it has been disabled.)", "votes": 14 }, { "id": 3168594, "postDate": "2025-04-02T15:28:46.220Z", "content": "In https://www.kaggle.com/code/dalloliogm/basic-prediction-using-cnn-and-image-patchesnotebook, I have written an utility function which shows the location of the spots, a square patch around them, and the distribution of cell types in it.
\nYou can see from this specific example, that there are patches that map to blank areas, yet still have a population of cell types. If the coordinates of the slices were correct, we would expect them to see zero abundance for all cell types.
\n
@gpch2159 were you able to understand what this misalignment mean, and how it might effect the training process?
", "rawMarkdown": "@gpch2159 were you able to understand what this misalignment mean, and how it might effect the training process?", "votes": 1, "replies": [ { "id": 3167989, "postDate": "2025-04-02T02:15:45.503Z", "content": "I have no further comments on it unless the host provides further info, because I am not an expert in this field.
", "rawMarkdown": "I have no further comments on it unless the host provides further info, because I am not an expert in this field.", "votes": 1 } ] }, { "id": 3162079, "postDate": "2025-03-28T19:18:54.967Z", "content": "@elucidata2025
\nPlease take a look at provided dataset, almost every slide has points outside the image where S_1 & S_2 have more points that are out of image. This will affect the model training.
@elucidata2025 Did you have a chance to look at the misalignment?
", "rawMarkdown": "@elucidata2025 Did you have a chance to look at the misalignment?", "votes": 1, "replies": [ { "id": 3164463, "postDate": "2025-03-31T18:24:43.647Z", "rawMarkdown": "", "isDeleted": true } ] }, { "id": 3159459, "postDate": "2025-03-25T16:15:13.207Z", "content": "That's a good point. Are you referring to the dots in the South part of S_2, for example?
\nIt looks like the whole image is misaligned, as the shape do not overlap perfectly. Is this what you mean?
I've created an EDA notebook to plot the abundance by Cell Type. If you scroll to the bottom, there are plots that show the total abundance in every slice. If you look closely at S_2, for example, there are dots that fall outside of the slice, yet they have high total abundance, meaning that there are some cells in there.
", "rawMarkdown": "That's a good point. Are you referring to the dots in the South part of S_2, for example? \nIt looks like the whole image is misaligned, as the shape do not overlap perfectly. Is this what you mean?\n\nI've created an[ EDA notebook](https://www.kaggle.com/code/dalloliogm/eda-plotting-cell-distribution-by-slice) to plot the abundance by Cell Type. If you scroll to the bottom, there are plots that show the total abundance in every slice. If you look closely at S_2, for example, there are dots that fall outside of the slice, yet they have high total abundance, meaning that there are some cells in there. ", "votes": 2, "replies": [ { "id": 3159778, "postDate": "2025-03-25T22:30:39.687Z", "content": "See these examples. The colours represent the total abundance across all cell types: violet dots have low overall abundance (e.g. not many cell types), while the other colours indicate higher abundance. There are a few spots, at the edges of the slides, that do not seem to be properly aligned, and that have high abundance.
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F650%2F7a054af9fc8b9a82f972f4ccfc7500d2%2F__results___11_1.png?generation=1743448529850771&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F650%2Fc7e35b57b91331f98b47c7b622052463%2F__results___10_1.png?generation=1743448541751895&alt=media\")
So can we just move the total spots so that is is properly aligned.
", "rawMarkdown": "So can we just move the total spots so that is is properly aligned. ", "votes": 1 }, { "id": 3160353, "postDate": "2025-03-26T16:53:32.950Z", "content": "That's a very good question. It is important to ensure that the spots are correctly aligned to the image - otherwise, it will be difficult to make predictions for the correct location. Unfortunately I do not know the answer.
\nIn my EDA notebook I use the snippet:
show_image True:\n ax.imshow(, aspect=)\n :\n # Manually set the axis limits to match the \n , = .shape[:]\n ax.set_xlim(, )\n ax.set_ylim(, ) # y-axis to mimic \n
\nbut as you can see, it doesn't perfectly align the spots to the image.
For S_1 x-50,y-60 gives a better alignment, similarly for S_2 x-95, y-55, Though these are just my assumptions.
\nAfter the changes in x and y axis, Do we need to use them for training, as the original loc is no longer true.
In https://www.kaggle.com/code/dalloliogm/basic-prediction-using-cnn-and-image-patchesnotebook, I have written an utility function which shows the location of the spots, a square patch around them, and the distribution of cell types in it.
\nYou can see from this specific example, that there are patches that map to blank areas, yet still have a population of cell types. If the coordinates of the slices were correct, we would expect them to see zero abundance for all cell types.
\n
@gpch2159 were you able to understand what this misalignment mean, and how it might effect the training process?
", "votes": 1, "replies": [ { "id": 3167989, "author_name": "GPCH", "author_url": "", "post_date": "2025-04-02T02:15:45.503000", "content": "I have no further comments on it unless the host provides further info, because I am not an expert in this field.
", "votes": 1, "replies": [] } ] }, { "id": 3162079, "author_name": "Ayush Raj", "author_url": "", "post_date": "2025-03-28T19:18:54.967000", "content": "@elucidata2025
\nPlease take a look at provided dataset, almost every slide has points outside the image where S_1 & S_2 have more points that are out of image. This will affect the model training.
@elucidata2025 Did you have a chance to look at the misalignment?
", "votes": 1, "replies": [ { "id": 3164463, "author_name": "", "author_url": "", "post_date": "2025-03-31T18:24:43.647000", "content": "", "votes": 0, "replies": [] } ] }, { "id": 3159459, "author_name": "Giovanni Marco Dall'Olio", "author_url": "", "post_date": "2025-03-25T16:15:13.207000", "content": "That's a good point. Are you referring to the dots in the South part of S_2, for example?
\nIt looks like the whole image is misaligned, as the shape do not overlap perfectly. Is this what you mean?
I've created an EDA notebook to plot the abundance by Cell Type. If you scroll to the bottom, there are plots that show the total abundance in every slice. If you look closely at S_2, for example, there are dots that fall outside of the slice, yet they have high total abundance, meaning that there are some cells in there.
", "votes": 2, "replies": [ { "id": 3159778, "author_name": "Giovanni Marco Dall'Olio", "author_url": "", "post_date": "2025-03-25T22:30:39.687000", "content": "See these examples. The colours represent the total abundance across all cell types: violet dots have low overall abundance (e.g. not many cell types), while the other colours indicate higher abundance. There are a few spots, at the edges of the slides, that do not seem to be properly aligned, and that have high abundance.
\n
\n
So can we just move the total spots so that is is properly aligned.
", "votes": 1, "replies": [] }, { "id": 3160353, "author_name": "Giovanni Marco Dall'Olio", "author_url": "", "post_date": "2025-03-26T16:53:32.950000", "content": "That's a very good question. It is important to ensure that the spots are correctly aligned to the image - otherwise, it will be difficult to make predictions for the correct location. Unfortunately I do not know the answer.
\nIn my EDA notebook I use the snippet:
show_image True:\n ax.imshow(, aspect=)\n :\n # Manually set the axis limits to match the \n , = .shape[:]\n ax.set_xlim(, )\n ax.set_ylim(, ) # y-axis to mimic \n
\nbut as you can see, it doesn't perfectly align the spots to the image.
For S_1 x-50,y-60 gives a better alignment, similarly for S_2 x-95, y-55, Though these are just my assumptions.
\nAfter the changes in x and y axis, Do we need to use them for training, as the original loc is no longer true.