![\"\"](\"https://i.ibb.co/f1dv3vG/Selection-207.png\")
\n
![\"\"](\"https://i.ibb.co/2sMjtcg/Selection-208.png\")
\n
\n
Instead of using segmentation images, I used box-annotated images.
\nI trained a classification model with box-annotated images, and its AUC validation score was about 0.9 (good result).
The idea was the radiologists might base on the presence of opacities to define one accurate from the four labels. In other words, there might be some links between four-label classification and opacity localization in this competition, thereby we cant do it in a separate way.
\nThen my simple stage was:
\nBut it didn't work on LB scores. 😪
", "rawMarkdown": "Instead of using segmentation images, I used box-annotated images. \nI trained a classification model with box-annotated images, and its AUC validation score was about 0.9 (good result).\n\nThe idea was the radiologists might base on the presence of opacities to define one accurate from the four labels. In other words, there might be some links between four-label classification and opacity localization in this competition, thereby we cant do it in a separate way.\n\nThen my simple stage was:\n1. Classification with normal images (model A)\n2. Bounding box detection model (YoloV5) --> box-annotated images --> classification model trained with box-annotated images (model B).\nMy result was the weighted ensemble of (model A, model B) + box detection model (YoloV5).\n\nBut it didn't work on LB scores. 😪", "votes": 9, "replies": [ { "id": 1344535, "postDate": "2021-06-11T02:06:40.650Z", "content": "My next plan is to build a multi-task learning model for both two tasks.
", "rawMarkdown": "My next plan is to build a multi-task learning model for both two tasks." }, { "id": 1344594, "postDate": "2021-06-11T03:24:48.680Z", "content": "Can you share what's your Yolov5 box ap ? I also trained an object detector using Yolov5 but the performance is not good(around 17 AP@IOU=0.5:0.95 and 45 AP@IOU=0.5)
", "rawMarkdown": "Can you share what's your Yolov5 box ap ? I also trained an object detector using Yolov5 but the performance is not good(around 17 AP@IOU=0.5:0.95 and 45 AP@IOU=0.5)" }, { "id": 1344629, "postDate": "2021-06-11T04:12:16.760Z", "content": "I haven't searched for the best hyperparameters in YoloV5, and my scores were exactly like that of yours.
", "rawMarkdown": "I haven't searched for the best hyperparameters in YoloV5, and my scores were exactly like that of yours." }, { "id": 1344800, "postDate": "2021-06-11T06:33:20.370Z", "content": "Hi @solosquad1999
\nI didn't get how will you use that box-annotated image for classification
@mrinath In the function get_item of class Dataset, I used cv2.rectangle to draw bounding boxes to images before fitting them into the models.
\nSorry for the late reply!
I still have some doubts, Do you mean that you only predict the image inside the bounding box and not the whole image.
\nThank you
@mrinath : No, what I mean is that I predict one of the four labels of the whole images (512x512) with boxes drawn on them.
", "rawMarkdown": "@mrinath : No, what I mean is that I predict one of the four labels of the whole images (512x512) with boxes drawn on them.", "votes": 2 } ] }, { "id": 1349431, "postDate": "2021-06-14T18:56:27.540Z", "content": "Hi Heng. Any idea or paper about predicting the annotator bias in deep learning? A brief analysis of RICORD dataset turns out that experts disagree more often than unanimous agree, especial when the disease is at the \"swing state\" such as mild or indeterminate. Study label for one patient across different time point may fluctuate significantly even if I consider the majority vote. I believe this is the nature of chest image annotation. RICORD may be used as the scientific ground truth as most images have been triple examined. In comparison, this competition has only ONE annotator. Bias can be significant and label can be deviated from the \"scientific ground truth\". RICORD data is small (~1000) so it's hard to make it a \"teacher\" NN. I am out of idea besides incorporating high-confidence RICORD data into SIIM. How to make the model bias towards the annotator's flavor seems the key to win.
", "rawMarkdown": "Hi Heng. Any idea or paper about predicting the annotator bias in deep learning? A brief analysis of RICORD dataset turns out that experts **disagree more often** than unanimous agree, especial when the disease is at the \"swing state\" such as mild or indeterminate. Study label for one patient across different time point may fluctuate significantly even if I consider the majority vote. I believe this is the nature of chest image annotation. RICORD may be used as the **scientific ground truth** as most images have been triple examined. In comparison, this competition has only ONE annotator. Bias can be significant and label can be deviated from the \"scientific ground truth\". RICORD data is small (~1000) so it's hard to make it a \"teacher\" NN. I am out of idea besides incorporating high-confidence RICORD data into SIIM. How to make the model bias towards the annotator's flavor seems the key to win. " }, { "id": 1344573, "postDate": "2021-06-11T02:51:39.187Z", "content": "Did anyone try stacking multiple patients' images and applying 3D Convolution (+ LSTM) to learn features in a simultaneous way?
", "rawMarkdown": "Did anyone try stacking multiple patients' images and applying 3D Convolution (+ LSTM) to learn features in a simultaneous way?", "replies": [ { "id": 1344912, "postDate": "2021-06-11T07:47:00.927Z", "content": "Could you please explaın what could be the benefıt of LSTM here? I,m a bit confused as there is no sequence data to apply LSTM
", "rawMarkdown": "Could you please explaın what could be the benefıt of LSTM here? I,m a bit confused as there is no sequence data to apply LSTM " } ] }, { "id": 1344287, "postDate": "2021-06-10T19:16:10.763Z", "content": "Are there data for segmentation available as any external data?
", "rawMarkdown": "Are there data for segmentation available as any external data? ", "replies": [ { "id": 1344545, "postDate": "2021-06-11T02:23:25.303Z", "content": "I think the idea is to fill the bounding box annotations.
", "rawMarkdown": "I think the idea is to fill the bounding box annotations.", "votes": 1 }, { "id": 1344801, "postDate": "2021-06-11T06:34:28.433Z", "content": "Could you explain how could we use it while classifying
", "rawMarkdown": "Could you explain how could we use it while classifying" }, { "id": 1345067, "postDate": "2021-06-11T09:32:19.703Z", "content": "you can use the bounding box in this kaggle dataset.
\nif you refer to the pdf document that describe the dataset, you can see that there is actually segmentation label (and also other radiologist report, etc)
![\"\"](\"https://i.ibb.co/hgTpv3p/Selection-212.png\")
i haven't checked in details yet. but i would like to use these extra labels if i can
", "rawMarkdown": "you can use the bounding box in this kaggle dataset.\nif you refer to the pdf document that describe the dataset, you can see that there is actually segmentation label (and also other radiologist report, etc)\n\n\n\ni haven't checked in details yet. but i would like to use these extra labels if i can", "votes": 2 }, { "id": 1353048, "postDate": "2021-06-16T19:54:16.983Z", "content": "Can you share the link to this paper?
", "rawMarkdown": "Can you share the link to this paper?" }, { "id": 1364009, "postDate": "2021-06-24T14:33:02.857Z", "content": "Can you share the link to the pdf document? Thanks
", "rawMarkdown": "Can you share the link to the pdf document? Thanks" }, { "id": 1364249, "postDate": "2021-06-24T18:35:46.387Z", "content": "I think the segmentation is only available for a subset of 10 images only as per the paper's abstract:- https://arxiv.org/pdf/2006.01174.pdf
", "rawMarkdown": "I think the segmentation is only available for a subset of 10 images only as per the paper's abstract:- https://arxiv.org/pdf/2006.01174.pdf\n" } ] }, { "id": 1344164, "postDate": "2021-06-10T16:51:07.513Z", "content": "Are you using or planning to use unet for segmentation ?
", "rawMarkdown": "Are you using or planning to use unet for segmentation ?" }, { "id": 1344923, "postDate": "2021-06-11T07:53:24.387Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 1344526, "author_name": "The fearless", "author_url": "", "post_date": "2021-06-11T01:48:32.277000", "content": "Instead of using segmentation images, I used box-annotated images.
\nI trained a classification model with box-annotated images, and its AUC validation score was about 0.9 (good result).
The idea was the radiologists might base on the presence of opacities to define one accurate from the four labels. In other words, there might be some links between four-label classification and opacity localization in this competition, thereby we cant do it in a separate way.
\nThen my simple stage was:
\nBut it didn't work on LB scores. 😪
", "votes": 9, "replies": [ { "id": 1344535, "author_name": "The fearless", "author_url": "", "post_date": "2021-06-11T02:06:40.650000", "content": "My next plan is to build a multi-task learning model for both two tasks.
", "votes": 0, "replies": [] }, { "id": 1344594, "author_name": "Anii", "author_url": "", "post_date": "2021-06-11T03:24:48.680000", "content": "Can you share what's your Yolov5 box ap ? I also trained an object detector using Yolov5 but the performance is not good(around 17 AP@IOU=0.5:0.95 and 45 AP@IOU=0.5)
", "votes": 0, "replies": [] }, { "id": 1344629, "author_name": "The fearless", "author_url": "", "post_date": "2021-06-11T04:12:16.760000", "content": "I haven't searched for the best hyperparameters in YoloV5, and my scores were exactly like that of yours.
", "votes": 0, "replies": [] }, { "id": 1344800, "author_name": "DeepUnderstanding", "author_url": "", "post_date": "2021-06-11T06:33:20.370000", "content": "Hi @solosquad1999
\nI didn't get how will you use that box-annotated image for classification
@mrinath In the function get_item of class Dataset, I used cv2.rectangle to draw bounding boxes to images before fitting them into the models.
\nSorry for the late reply!
I still have some doubts, Do you mean that you only predict the image inside the bounding box and not the whole image.
\nThank you
@mrinath : No, what I mean is that I predict one of the four labels of the whole images (512x512) with boxes drawn on them.
", "votes": 2, "replies": [] } ] }, { "id": 1349431, "author_name": "human intelligence", "author_url": "", "post_date": "2021-06-14T18:56:27.540000", "content": "Hi Heng. Any idea or paper about predicting the annotator bias in deep learning? A brief analysis of RICORD dataset turns out that experts disagree more often than unanimous agree, especial when the disease is at the \"swing state\" such as mild or indeterminate. Study label for one patient across different time point may fluctuate significantly even if I consider the majority vote. I believe this is the nature of chest image annotation. RICORD may be used as the scientific ground truth as most images have been triple examined. In comparison, this competition has only ONE annotator. Bias can be significant and label can be deviated from the \"scientific ground truth\". RICORD data is small (~1000) so it's hard to make it a \"teacher\" NN. I am out of idea besides incorporating high-confidence RICORD data into SIIM. How to make the model bias towards the annotator's flavor seems the key to win.
", "votes": 0, "replies": [] }, { "id": 1344573, "author_name": "The fearless", "author_url": "", "post_date": "2021-06-11T02:51:39.187000", "content": "Did anyone try stacking multiple patients' images and applying 3D Convolution (+ LSTM) to learn features in a simultaneous way?
", "votes": 0, "replies": [ { "id": 1344912, "author_name": "erkut", "author_url": "", "post_date": "2021-06-11T07:47:00.927000", "content": "Could you please explaın what could be the benefıt of LSTM here? I,m a bit confused as there is no sequence data to apply LSTM
", "votes": 0, "replies": [] } ] }, { "id": 1344287, "author_name": "DeepUnderstanding", "author_url": "", "post_date": "2021-06-10T19:16:10.763000", "content": "Are there data for segmentation available as any external data?
", "votes": 0, "replies": [ { "id": 1344545, "author_name": "Kerem Turgutlu", "author_url": "", "post_date": "2021-06-11T02:23:25.303000", "content": "I think the idea is to fill the bounding box annotations.
", "votes": 1, "replies": [] }, { "id": 1344801, "author_name": "DeepUnderstanding", "author_url": "", "post_date": "2021-06-11T06:34:28.433000", "content": "Could you explain how could we use it while classifying
", "votes": 0, "replies": [] }, { "id": 1345067, "author_name": "hengck23", "author_url": "", "post_date": "2021-06-11T09:32:19.703000", "content": "you can use the bounding box in this kaggle dataset.
\nif you refer to the pdf document that describe the dataset, you can see that there is actually segmentation label (and also other radiologist report, etc)
i haven't checked in details yet. but i would like to use these extra labels if i can
", "votes": 2, "replies": [] }, { "id": 1353048, "author_name": "William Green", "author_url": "", "post_date": "2021-06-16T19:54:16.983000", "content": "Can you share the link to this paper?
", "votes": 0, "replies": [] }, { "id": 1364009, "author_name": "Wang Xinliang", "author_url": "", "post_date": "2021-06-24T14:33:02.857000", "content": "Can you share the link to the pdf document? Thanks
", "votes": 0, "replies": [] }, { "id": 1364249, "author_name": "Harsh Abhishek", "author_url": "", "post_date": "2021-06-24T18:35:46.387000", "content": "I think the segmentation is only available for a subset of 10 images only as per the paper's abstract:- https://arxiv.org/pdf/2006.01174.pdf
", "votes": 0, "replies": [] } ] }, { "id": 1344164, "author_name": "Varun Dutt", "author_url": "", "post_date": "2021-06-10T16:51:07.513000", "content": "Are you using or planning to use unet for segmentation ?
", "votes": 0, "replies": [] }, { "id": 1344923, "author_name": "", "author_url": "", "post_date": "2021-06-11T07:53:24.387000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1344056": "\n", "1344526": "Instead of using segmentation images, I used box-annotated images. \nI trained a classification model with box-annotated images, and its AUC validation score was about 0.9 (good result).\n\nThe idea was the radiologists might base on the presence of opacities to define one accurate from the four labels. In other words, there might be some links between four-label classification and opacity localization in this competition, thereby we cant do it in a separate way.\n\nThen my simple stage was:\n1. Classification with normal images (model A)\n2. Bounding box detection model (YoloV5) --> box-annotated images --> classification model trained with box-annotated images (model B).\nMy result was the weighted ensemble of (model A, model B) + box detection model (YoloV5).\n\nBut it didn't work on LB scores. 😪", "1349431": "Hi Heng. Any idea or paper about predicting the annotator bias in deep learning? A brief analysis of RICORD dataset turns out that experts **disagree more often** than unanimous agree, especial when the disease is at the \"swing state\" such as mild or indeterminate. Study label for one patient across different time point may fluctuate significantly even if I consider the majority vote. I believe this is the nature of chest image annotation. RICORD may be used as the **scientific ground truth** as most images have been triple examined. In comparison, this competition has only ONE annotator. Bias can be significant and label can be deviated from the \"scientific ground truth\". RICORD data is small (~1000) so it's hard to make it a \"teacher\" NN. I am out of idea besides incorporating high-confidence RICORD data into SIIM. How to make the model bias towards the annotator's flavor seems the key to win. ", "1344573": "Did anyone try stacking multiple patients' images and applying 3D Convolution (+ LSTM) to learn features in a simultaneous way?", "1344287": "Are there data for segmentation available as any external data? ", "1344164": "Are you using or planning to use unet for segmentation ?", "1344923": "" } }