![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2523201%2F4e4b4200e9f2986f5b5bae25128a5537%2F2020-12-20%2013-35-40.png?generation=1608438956908406&alt=media\")
I participated in this competition and read a paper.
\nSo, in such models, we have to do various tasks such as evaluating the presence, classification, and location of catheters.
\nI suggest splitting the model. Specifically, I propose to insert a segmentation model into the training sequence. The trained model is then used to visualize the catheter's location in the image, reducing the burden on the CNN model.
\nI should add that I am new to deep learning, so I don't know how these ideas work in detail.
\nPlease share your thoughts on this idea with me.
(https://arxiv.org/pdf/1907.01656.pdf)
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2523201%2F50ba78e606577292b9abe054ee9c68a4%2F2020-12-20%2013-20-42.png?generation=1608438077434281&alt=media\")
I participated in this competition and read a paper.
\nSo, in such models, we have to do various tasks such as evaluating the presence, classification, and location of catheters.
\nI suggest splitting the model. Specifically, I propose to insert a segmentation model into the training sequence. The trained model is then used to visualize the catheter's location in the image, reducing the burden on the CNN model.
\nI should add that I am new to deep learning, so I don't know how these ideas work in detail.
\nPlease share your thoughts on this idea with me.
(https://arxiv.org/pdf/1907.01656.pdf)
\n
Hey @chizuchizu thanks for suggesting the paper.π
\nAlso the end-to-end deep learning models (without segmentation) are getting the AUC: 0.93 on validation data.
\nIts nice idea to put segmentation model into training phase and then train CNN classifier on top of it, on the other hand it's very time consuming since the end-to-end deep CNN model is taking around 20-40 min for single epoch.
Hi, @akhileshdkapse
\nThank you for sharing your opinion with us.
I see. It's true that the CNN model itself takes a long time, so if you add a segmentation model to handle even more difficult tasks, it will take an enormous amount of time.
\nI have a few suggestions based on this opinion.
\nIf we were to adopt the third idea, it would be multimodal learning (CNNs that everyone uses plus classification CNNs).
\nIt's a long story, but I'd like to know what you think about this.
", "rawMarkdown": "Hi, @akhileshdkapse \nThank you for sharing your opinion with us.\n\nI see. It's true that the CNN model itself takes a long time, so if you add a segmentation model to handle even more difficult tasks, it will take an enormous amount of time.\n\nI have a few suggestions based on this opinion.\n\n- Predict the end of the catheter. (Only the lower right coordinate of the catheter rectangle)\n- Predict the rectangle of the catheter (object detection model, but I'm sure this is also a heavy process)\n- A classification model to predict the type of catheter present.\n\nIf we were to adopt the third idea, it would be multimodal learning (CNNs that everyone uses plus classification CNNs).\n\nIt's a long story, but I'd like to know what you think about this.", "votes": null }, { "id": "1119932", "postDate": "12/20/2020 13:11:09", "content": "For now I have 3 more suggestions, and I will add more timely-
\nUsing 1st one I achieved 0.933~0.944 Scoreπ.
\nHope, more member will share different approaches π
", "rawMarkdown": "For now I have 3 more suggestions, and I will add more timely-\n1. Train End-to-End CNN model.\n2. K-Fold in Training data can also help accordingly. \n3. Considering this topic- https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/205243\n\nUsing 1st one I achieved 0.933~0.944 Scoreπ.\n\nHope, more member will share different approaches π", "votes": null }, { "id": "1120028", "postDate": "12/20/2020 14:58:28", "content": "I like the idea of using segmentation for this task, thanks for share.
", "rawMarkdown": "I like the idea of using **segmentation** for this task, thanks for share.", "votes": null }, { "id": "1120033", "postDate": "12/20/2020 15:01:18", "content": "while segmentation is possible, e.g.
\none has to take note:
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F64cb0ff91c195252037dfe09645e2ae1%2FSelection_104.png?generation=1608476474325816&alt=media\")
classification is already looking at the tip of the tube as shown the the CAM activation map for identification of ETT.
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fc4e246dcabc2b0a8957752be62153881%2FSelection_070.png?generation=1608476637584614&alt=media\")
@hengck23 special thanks for sharing man!
\nCan you please help in this topic-
\nhttps://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/205144
@hengck23 Thank you for your comment.
\nYour opinion is correct.
\nI need to use an image with a background included. (I am close to your idea)
\nFor example, if we use just a CNN model to predict whether a tube is present or not, we might be able to reduce the prediction from 0.1 to 0. (If the model is reliable.)
\nI'll keep trying!!!! Let's go!!!!!!
", "rawMarkdown": "hengck23 Thank you for your comment.\n\nYour opinion is correct.\n\nI need to use an image with a background included. (I am close to [your idea](https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/205243))\n\nFor example, if we use just a CNN model to predict whether a tube is present or not, we might be able to reduce the prediction from 0.1 to 0. (If the model is reliable.)\n\n\nI'll keep trying!!!! Let's go!!!!!!", "votes": null }, { "id": "1121125", "postDate": "12/21/2020 11:48:23", "content": "Thanks for replying and sharing. I'm going to try a few ideas.
", "rawMarkdown": "Thanks for replying and sharing. I'm going to try a few ideas.", "votes": null }, { "id": "1122416", "postDate": "12/22/2020 12:30:39", "content": "[update]
\nI created a classification model that predicts the presence of a catheter, separate from the main model, with an AUC score of 0.98.
\nUsing that model, I set predictions below a certain threshold to zero (labeled catheters that would not be present). However, the score dropped significantly to 0.949->0.936.
\nI would also like to try multimodal learning.
", "rawMarkdown": "[update]\n\nI created a classification model that predicts the presence of a catheter, separate from the main model, with an AUC score of 0.98.\n\nUsing that model, I set predictions below a certain threshold to zero (labeled catheters that would not be present). However, the score dropped significantly to 0.949->0.936.\n\nI would also like to try multimodal learning.", "votes": null }, { "id": "1138961", "postDate": "01/05/2021 04:49:08", "content": "Hi there,
\nHave you had much success with multimodal learning? I found this paper very interesting, but I can't work out how to implement the 'feature extraction' section with e.g. a ResNet.
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4271956%2Fc804f27442bced443c39489ed4e80344%2Finbox_2523201_50ba78e606577292b9abe054ee9c68a4_2020-12-20%2013-20-42.png?generation=1609822120690182&alt=media\")
It's not done.\" I understand \"feature extraction\" as simply extracting features of an image across a CNN model.
", "rawMarkdown": "It's not done.\" I understand \"feature extraction\" as simply extracting features of an image across a CNN model.", "votes": null }, { "id": "1183487", "postDate": "02/03/2021 03:32:01", "content": "Thanks for sharing CAM result, if we can get right CAM with only classification supervision,i think the number of train images are enough,so use tube annotation can't get too much improvement.
", "rawMarkdown": "Thanks for sharing CAM result, if we can get right CAM with only classification supervision,i think the number of train images are enough,so use tube annotation can't get too much improvement.", "votes": null }, { "id": "1183491", "postDate": "02/03/2021 03:37:54", "content": "Yes,predicting the end of the catheter maybe an efficient way,but we only get classification annotation,how to get the position annotation to predict the position of the end of the catheter?
", "rawMarkdown": "Yes,predicting the end of the catheter maybe an efficient way,but we only get classification annotation,how to get the position annotation to predict the position of the end of the catheter?", "votes": null }, { "id": "1183926", "postDate": "02/03/2021 09:55:05", "content": "I thought I could use both ends of the catheter when I represented it as a line.
\nBut I don't think it's likely to work.
Yes,I think so.Maybe we need to think other methods to use classification annotation efficiently.I think that using only classification labels, the accuracy of the model still has a lot of room for improvement.
", "rawMarkdown": "Yes,I think so.Maybe we need to think other methods to use classification annotation efficiently.I think that using only classification labels, the accuracy of the model still has a lot of room for improvement.", "votes": null } ], "comments": [ { "id": 1119686, "author_name": "akhileshdkapse", "author_url": "", "post_date": "12/20/2020 09:44:15", "content": "Hey @chizuchizu thanks for suggesting the paper.π
\nAlso the end-to-end deep learning models (without segmentation) are getting the AUC: 0.93 on validation data.
\nIts nice idea to put segmentation model into training phase and then train CNN classifier on top of it, on the other hand it's very time consuming since the end-to-end deep CNN model is taking around 20-40 min for single epoch.
Hi, @akhileshdkapse
\nThank you for sharing your opinion with us.
I see. It's true that the CNN model itself takes a long time, so if you add a segmentation model to handle even more difficult tasks, it will take an enormous amount of time.
\nI have a few suggestions based on this opinion.
\nIf we were to adopt the third idea, it would be multimodal learning (CNNs that everyone uses plus classification CNNs).
\nIt's a long story, but I'd like to know what you think about this.
", "votes": null, "replies": [] }, { "id": 1119932, "author_name": "akhileshdkapse", "author_url": "", "post_date": "12/20/2020 13:11:09", "content": "For now I have 3 more suggestions, and I will add more timely-
\nUsing 1st one I achieved 0.933~0.944 Scoreπ.
\nHope, more member will share different approaches π
", "votes": null, "replies": [] }, { "id": 1121125, "author_name": "chizuchizu", "author_url": "", "post_date": "12/21/2020 11:48:23", "content": "Thanks for replying and sharing. I'm going to try a few ideas.
", "votes": null, "replies": [] }, { "id": 1183491, "author_name": "jianxinhu", "author_url": "", "post_date": "02/03/2021 03:37:54", "content": "Yes,predicting the end of the catheter maybe an efficient way,but we only get classification annotation,how to get the position annotation to predict the position of the end of the catheter?
", "votes": null, "replies": [] }, { "id": 1183926, "author_name": "chizuchizu", "author_url": "", "post_date": "02/03/2021 09:55:05", "content": "I thought I could use both ends of the catheter when I represented it as a line.
\nBut I don't think it's likely to work.
Yes,I think so.Maybe we need to think other methods to use classification annotation efficiently.I think that using only classification labels, the accuracy of the model still has a lot of room for improvement.
", "votes": null, "replies": [] } ] }, { "id": 1120028, "author_name": "hiramcho", "author_url": "", "post_date": "12/20/2020 14:58:28", "content": "I like the idea of using segmentation for this task, thanks for share.
", "votes": null, "replies": [] }, { "id": 1120033, "author_name": "hengck23", "author_url": "", "post_date": "12/20/2020 15:01:18", "content": "while segmentation is possible, e.g.
\none has to take note:
\n
classification is already looking at the tip of the tube as shown the the CAM activation map for identification of ETT.
\n
@hengck23 special thanks for sharing man!
\nCan you please help in this topic-
\nhttps://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/205144
@hengck23 Thank you for your comment.
\nYour opinion is correct.
\nI need to use an image with a background included. (I am close to your idea)
\nFor example, if we use just a CNN model to predict whether a tube is present or not, we might be able to reduce the prediction from 0.1 to 0. (If the model is reliable.)
\nI'll keep trying!!!! Let's go!!!!!!
", "votes": null, "replies": [] }, { "id": 1183487, "author_name": "jianxinhu", "author_url": "", "post_date": "02/03/2021 03:32:01", "content": "Thanks for sharing CAM result, if we can get right CAM with only classification supervision,i think the number of train images are enough,so use tube annotation can't get too much improvement.
", "votes": null, "replies": [] } ] }, { "id": 1122416, "author_name": "chizuchizu", "author_url": "", "post_date": "12/22/2020 12:30:39", "content": "[update]
\nI created a classification model that predicts the presence of a catheter, separate from the main model, with an AUC score of 0.98.
\nUsing that model, I set predictions below a certain threshold to zero (labeled catheters that would not be present). However, the score dropped significantly to 0.949->0.936.
\nI would also like to try multimodal learning.
", "votes": null, "replies": [ { "id": 1138961, "author_name": "reubenschmidt", "author_url": "", "post_date": "01/05/2021 04:49:08", "content": "Hi there,
\nHave you had much success with multimodal learning? I found this paper very interesting, but I can't work out how to implement the 'feature extraction' section with e.g. a ResNet.
It's not done.\" I understand \"feature extraction\" as simply extracting features of an image across a CNN model.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1119476": "I participated in this competition and read a [paper](https://arxiv.org/pdf/1907.01656.pdf).\n\nSo, in such models, we have to do various tasks such as evaluating the presence, classification, and location of catheters.\n\nI suggest splitting the model. Specifically, I propose to insert a segmentation model into the training sequence. The trained model is then used to visualize the catheter's location in the image, reducing the burden on the CNN model.\n\nI should add that I am new to deep learning, so I don't know how these ideas work in detail.\nPlease share your thoughts on this idea with me.\n\n\n\n(https://arxiv.org/pdf/1907.01656.pdf)\n\n", "1119686": "Hey @chizuchizu thanks for suggesting the paper.π\nAlso the end-to-end deep learning models (without segmentation) are getting the AUC: 0.93 on validation data.\nIts nice idea to put segmentation model into training phase and then train CNN classifier on top of it, on the other hand it's very time consuming since the end-to-end deep CNN model is taking around 20-40 min for single epoch.", "1119716": "Hi, @akhileshdkapse \nThank you for sharing your opinion with us.\n\nI see. It's true that the CNN model itself takes a long time, so if you add a segmentation model to handle even more difficult tasks, it will take an enormous amount of time.\n\nI have a few suggestions based on this opinion.\n\n- Predict the end of the catheter. (Only the lower right coordinate of the catheter rectangle)\n- Predict the rectangle of the catheter (object detection model, but I'm sure this is also a heavy process)\n- A classification model to predict the type of catheter present.\n\nIf we were to adopt the third idea, it would be multimodal learning (CNNs that everyone uses plus classification CNNs).\n\nIt's a long story, but I'd like to know what you think about this.", "1119932": "For now I have 3 more suggestions, and I will add more timely-\n1. Train End-to-End CNN model.\n2. K-Fold in Training data can also help accordingly. \n3. Considering this topic- https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/205243\n\nUsing 1st one I achieved 0.933~0.944 Scoreπ.\n\nHope, more member will share different approaches π", "1120028": "I like the idea of using **segmentation** for this task, thanks for share.", "1120033": "while segmentation is possible, e.g.\n- as auxiliary task loss\n- as input to classification, etc ...\n\none has to take note:\n\n", "1120041": "classification is already looking at the tip of the tube as shown the the CAM activation map for identification of ETT.\n\n", "1120080": "hengck23 special thanks for sharing man! \nCan you please help in this topic-\nhttps://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/205144", "1121124": "hengck23 Thank you for your comment.\n\nYour opinion is correct.\n\nI need to use an image with a background included. (I am close to [your idea](https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/205243))\n\nFor example, if we use just a CNN model to predict whether a tube is present or not, we might be able to reduce the prediction from 0.1 to 0. (If the model is reliable.)\n\n\nI'll keep trying!!!! Let's go!!!!!!", "1121125": "Thanks for replying and sharing. I'm going to try a few ideas.", "1122416": "[update]\n\nI created a classification model that predicts the presence of a catheter, separate from the main model, with an AUC score of 0.98.\n\nUsing that model, I set predictions below a certain threshold to zero (labeled catheters that would not be present). However, the score dropped significantly to 0.949->0.936.\n\nI would also like to try multimodal learning.", "1138961": "Hi there,\nHave you had much success with multimodal learning? I found this paper very interesting, but I can't work out how to implement the 'feature extraction' section with e.g. a ResNet.\n\n", "1139045": "It's not done.\" I understand \"feature extraction\" as simply extracting features of an image across a CNN model.", "1183487": "Thanks for sharing CAM result, if we can get right CAM with only classification supervision,i think the number of train images are enough,so use tube annotation can't get too much improvement.", "1183491": "Yes,predicting the end of the catheter maybe an efficient way,but we only get classification annotation,how to get the position annotation to predict the position of the end of the catheter?", "1183926": "I thought I could use both ends of the catheter when I represented it as a line.\nBut I don't think it's likely to work.", "1184128": "Yes,I think so.Maybe we need to think other methods to use classification annotation efficiently.I think that using only classification labels, the accuracy of the model still has a lot of room for improvement." }, "source": "meta" }