Hello everyone!!
\nSo far I've been reading excellent discussion and notebooks such as;
\nhttps://www.kaggle.com/c/hpa-single-cell-image-classification/discussion/214659
\nhttps://www.kaggle.com/c/hpa-single-cell-image-classification/discussion/215141
\nhttps://www.kaggle.com/lnhtrang/single-cell-patterns
\nhttps://www.kaggle.com/thedrcat/hpa-single-cell-classification-eda
But still I'm confused at instance segmentation's metric.
\nFor example, hosts say that we should segment the full cell, but if we use the model segmenting the whole part of every cell for prediction, there's gonna be many FP (IoU < 0.6) when true mask locates only inside the nucleus (ex: 5. nuclear bodies).
\nHere is my question.
\nShould we segment each cell efficiently in order to get high score, like only nucleus, plasma membrane or cytoplasm etc?
\nOr just segmenting the full cell is inevitable approach?
\nI guess it's a little bit risky to use different segmentation models for each cell, though…
Hello everyone!!
\nSo far I've been reading excellent discussion and notebooks such as;
\nhttps://www.kaggle.com/c/hpa-single-cell-image-classification/discussion/214659
\nhttps://www.kaggle.com/c/hpa-single-cell-image-classification/discussion/215141
\nhttps://www.kaggle.com/lnhtrang/single-cell-patterns
\nhttps://www.kaggle.com/thedrcat/hpa-single-cell-classification-eda
But still I'm confused at instance segmentation's metric.
\nFor example, hosts say that we should segment the full cell, but if we use the model segmenting the whole part of every cell for prediction, there's gonna be many FP (IoU < 0.6) when true mask locates only inside the nucleus (ex: 5. nuclear bodies).
\nHere is my question.
\nShould we segment each cell efficiently in order to get high score, like only nucleus, plasma membrane or cytoplasm etc?
\nOr just segmenting the full cell is inevitable approach?
\nI guess it's a little bit risky to use different segmentation models for each cell, though…
In this competition, you should submit the whole cell segmentation.
\nFrom a research perspective, individual organelle segmentation is definitely an interesting task. But since they are morphologically so different as you see in the Single Cell Pattern notebook, and so are their functions, it is more efficient to develop and analyze a model for each organelle of interest. In fact this is what happens in the biological image analysis space right now with individual deep learning models for Mitochondria, Golgi etc.
\nThere are more considerations that we took into account when designing this challenge. We want to study the integrated cell, with organelles in relation to it. Cell segmentation task can be considered an auxiliary task (depending on your approach to modelling of course) on a generalized deep learning model of the cell.
", "rawMarkdown": "In this competition, you should submit the whole cell segmentation.\n\nFrom a research perspective, individual organelle segmentation is definitely an interesting task. But since they are morphologically so different as you see in the Single Cell Pattern notebook, and so are their functions, it is more efficient to develop and analyze a model for each organelle of interest. In fact this is what happens in the biological image analysis space right now with individual deep learning models for Mitochondria, Golgi etc. \n\nThere are more considerations that we took into account when designing this challenge. We want to study the integrated cell, with organelles in relation to it. Cell segmentation task can be considered an auxiliary task (depending on your approach to modelling of course) on a generalized deep learning model of the cell. ", "votes": 2, "replies": [ { "id": 1194714, "postDate": "2021-02-10T10:20:55.223Z", "content": "Thank you so much for the fast clarification.
\nNow I can concentrate on this interesting task.
This clears up the confusion, thanks a lot!
", "rawMarkdown": "This clears up the confusion, thanks a lot!", "votes": 1 }, { "id": 1207404, "postDate": "2021-02-17T20:34:12.983Z", "content": "Thus @Inhtrang, the main task is to correctly label each segmented cell? Segmentation is an auxiliary task.
\nAlso, will the scoring depend on labeling the cells that have the presence of the protein of interest? Are we to leave the cells that don't have any overlap with protein(green channel)?
", "rawMarkdown": "Thus @Inhtrang, the main task is to correctly label each segmented cell? Segmentation is an auxiliary task.\n\nAlso, will the scoring depend on labeling the cells that have the presence of the protein of interest? Are we to leave the cells that don't have any overlap with protein(green channel)?" }, { "id": 1208162, "postDate": "2021-02-18T07:32:12.253Z", "content": "Hi! The task is predicting subcellular patterns for each cell. Segmentation is a way to identify which cell you are predicting. It can also be considered as auxiliary task depending on your modelling approach.
\nScoring depends on classifying each cell correctly in one or more of the 19 classes. Please check out this Single Cell Pattern notebook to understand more about each pattern in this competition, including Negative class.
Got it! Thanks.
\nThus in the submission.csv file, the RLE encoded mask is just a way for the evaluation system to understand to which cell a label(s) is associated. Am I correct?
In this competition, you should submit the whole cell segmentation.
\nFrom a research perspective, individual organelle segmentation is definitely an interesting task. But since they are morphologically so different as you see in the Single Cell Pattern notebook, and so are their functions, it is more efficient to develop and analyze a model for each organelle of interest. In fact this is what happens in the biological image analysis space right now with individual deep learning models for Mitochondria, Golgi etc.
\nThere are more considerations that we took into account when designing this challenge. We want to study the integrated cell, with organelles in relation to it. Cell segmentation task can be considered an auxiliary task (depending on your approach to modelling of course) on a generalized deep learning model of the cell.
", "votes": 2, "replies": [ { "id": 1194714, "author_name": "cool_rabbit", "author_url": "", "post_date": "2021-02-10T10:20:55.223000", "content": "Thank you so much for the fast clarification.
\nNow I can concentrate on this interesting task.
This clears up the confusion, thanks a lot!
", "votes": 1, "replies": [] }, { "id": 1207404, "author_name": "Ayush Thakur", "author_url": "", "post_date": "2021-02-17T20:34:12.983000", "content": "Thus @Inhtrang, the main task is to correctly label each segmented cell? Segmentation is an auxiliary task.
\nAlso, will the scoring depend on labeling the cells that have the presence of the protein of interest? Are we to leave the cells that don't have any overlap with protein(green channel)?
", "votes": 0, "replies": [] }, { "id": 1208162, "author_name": "Trang Le", "author_url": "", "post_date": "2021-02-18T07:32:12.253000", "content": "Hi! The task is predicting subcellular patterns for each cell. Segmentation is a way to identify which cell you are predicting. It can also be considered as auxiliary task depending on your modelling approach.
\nScoring depends on classifying each cell correctly in one or more of the 19 classes. Please check out this Single Cell Pattern notebook to understand more about each pattern in this competition, including Negative class.
Got it! Thanks.
\nThus in the submission.csv file, the RLE encoded mask is just a way for the evaluation system to understand to which cell a label(s) is associated. Am I correct?