![\"ROI](\"https://i.ibb.co/LdrhsWn/br001.jpg\")
During the visual data analysis I noticed that there is a large variation in the arrangement of the object in the images. In addition, some objects occupy only a small part of the image. By converting (resizing) without ROI extraction we have a very inefficient use of the reduced image. Most of the picture is blank.
\nSolution:
\nResult on train DS:
\nModel performance (on my validation DS): mAP@50 -> 0.995, mAP50-95 -> 0.914
\n
Notebook: https://www.kaggle.com/code/remekkinas/breast-cancer-roi-brest-extractor
", "messages": [ { "id": 2051208, "postDate": "2022-12-01T09:28:18.863Z", "content": "During the visual data analysis I noticed that there is a large variation in the arrangement of the object in the images. In addition, some objects occupy only a small part of the image. By converting (resizing) without ROI extraction we have a very inefficient use of the reduced image. Most of the picture is blank.
\nSolution:
\nResult on train DS:
\nModel performance (on my validation DS): mAP@50 -> 0.995, mAP50-95 -> 0.914
\n
Notebook: https://www.kaggle.com/code/remekkinas/breast-cancer-roi-brest-extractor
", "rawMarkdown": "During the visual data analysis I noticed that there is a large variation in the arrangement of the object in the images. In addition, some objects occupy only a small part of the image. By converting (resizing) without ROI extraction we have a very inefficient use of the reduced image. Most of the picture is blank.\n\n**Solution:**\n* annotate data - I annotated about 500 images in a human in the loop technique (3 models were created - I started from 300 images and ended up about 500)\n* train object detector - I used yolov5 (small - balance between accuracy and speed)\n\n**Result on train DS:**\n* 54601 images processed successfully\n* 105 images - detection failed\n\nModel performance (on my validation DS): mAP@50 -> 0.995, mAP50-95 -> 0.914\n\n\n\nNotebook: https://www.kaggle.com/code/remekkinas/breast-cancer-roi-brest-extractor", "votes": 24 }, { "id": 2051338, "postDate": "2022-12-01T10:52:30.633Z", "content": "Thanks, @remekkinas for sharing. This is great!
", "rawMarkdown": "Thanks, @remekkinas for sharing. This is great!", "votes": 1, "replies": [ { "id": 2051346, "postDate": "2022-12-01T11:02:14.947Z", "content": "Thank you very much :)
", "rawMarkdown": "Thank you very much :) " } ] }, { "id": 2117043, "postDate": "2023-01-26T23:23:39.180Z", "content": "With respect, I believe this kind of task does not require the complexity of annotating and then training a model for it. It can be done using a simple function like this:
\ndef crop_func(img):
\n thresholded = cv2.threshold(np.array(img), 0, 255, cv2.THRESH_OTSU)
\n bbox = cv2.boundingRect(thresholded[1])
\n x, y, w, h = bbox
\n cr_img = img[y:y+h, x:x+w]
\n return cr_img
There are always outliers and a detection model is able to generalize much better (almost 100% coverage) from my tests.
", "rawMarkdown": "There are always outliers and a detection model is able to generalize much better (almost 100% coverage) from my tests." }, { "id": 2117352, "postDate": "2023-01-27T08:01:36.007Z", "content": "The way you presented is totally different.
\n![\"\"](\"https://i.ibb.co/d7YkpqF/roi.jpg\")
AI/ML (art of details)
", "rawMarkdown": "The way you presented is totally different. \n- you do not crop ROI - you crop contour\n- my way is to crop breast (center part)\n\n\n\nAI/ML (art of details)", "votes": 1 } ] }, { "id": 2053236, "postDate": "2022-12-03T02:01:54.177Z", "content": "Thanks for sharing, really good dataset. I seem to be getting a slightly lower CV score with ROI cropping. Did this work for anyone else? @remekkinas How did this go for you?
", "rawMarkdown": "Thanks for sharing, really good dataset. I seem to be getting a slightly lower CV score with ROI cropping. Did this work for anyone else? @remekkinas How did this go for you?", "replies": [ { "id": 2053255, "postDate": "2022-12-03T03:21:26.327Z", "content": "Ok, it started doing better at later epochs. Good stuff 👍
", "rawMarkdown": "Ok, it started doing better at later epochs. Good stuff 👍", "votes": 1 } ] } ], "comments": [ { "id": 2051338, "author_name": "Mohammad Dehghanmanshadi", "author_url": "", "post_date": "2022-12-01T10:52:30.633000", "content": "Thanks, @remekkinas for sharing. This is great!
", "votes": 1, "replies": [ { "id": 2051346, "author_name": "Remek Kinas", "author_url": "", "post_date": "2022-12-01T11:02:14.947000", "content": "Thank you very much :)
", "votes": 0, "replies": [] } ] }, { "id": 2117043, "author_name": "Sina Mehdinia", "author_url": "", "post_date": "2023-01-26T23:23:39.180000", "content": "With respect, I believe this kind of task does not require the complexity of annotating and then training a model for it. It can be done using a simple function like this:
\ndef crop_func(img):
\n thresholded = cv2.threshold(np.array(img), 0, 255, cv2.THRESH_OTSU)
\n bbox = cv2.boundingRect(thresholded[1])
\n x, y, w, h = bbox
\n cr_img = img[y:y+h, x:x+w]
\n return cr_img
There are always outliers and a detection model is able to generalize much better (almost 100% coverage) from my tests.
", "votes": 0, "replies": [] }, { "id": 2117352, "author_name": "Remek Kinas", "author_url": "", "post_date": "2023-01-27T08:01:36.007000", "content": "The way you presented is totally different.
\n
AI/ML (art of details)
", "votes": 1, "replies": [] } ] }, { "id": 2053236, "author_name": "Ivan Aerlic", "author_url": "", "post_date": "2022-12-03T02:01:54.177000", "content": "Thanks for sharing, really good dataset. I seem to be getting a slightly lower CV score with ROI cropping. Did this work for anyone else? @remekkinas How did this go for you?
", "votes": 0, "replies": [ { "id": 2053255, "author_name": "Ivan Aerlic", "author_url": "", "post_date": "2022-12-03T03:21:26.327000", "content": "Ok, it started doing better at later epochs. Good stuff 👍
", "votes": 1, "replies": [] } ] } ], "raw_markdown_by_id": { "2051208": "During the visual data analysis I noticed that there is a large variation in the arrangement of the object in the images. In addition, some objects occupy only a small part of the image. By converting (resizing) without ROI extraction we have a very inefficient use of the reduced image. Most of the picture is blank.\n\n**Solution:**\n* annotate data - I annotated about 500 images in a human in the loop technique (3 models were created - I started from 300 images and ended up about 500)\n* train object detector - I used yolov5 (small - balance between accuracy and speed)\n\n**Result on train DS:**\n* 54601 images processed successfully\n* 105 images - detection failed\n\nModel performance (on my validation DS): mAP@50 -> 0.995, mAP50-95 -> 0.914\n\n\n\nNotebook: https://www.kaggle.com/code/remekkinas/breast-cancer-roi-brest-extractor", "2051338": "Thanks, @remekkinas for sharing. This is great!", "2117043": "With respect, I believe this kind of task does not require the complexity of annotating and then training a model for it. It can be done using a simple function like this: \n\ndef crop_func(img):\n thresholded = cv2.threshold(np.array(img), 0, 255, cv2.THRESH_OTSU)\n bbox = cv2.boundingRect(thresholded[1])\n x, y, w, h = bbox\n cr_img = img[y:y+h, x:x+w]\n return cr_img", "2053236": "Thanks for sharing, really good dataset. I seem to be getting a slightly lower CV score with ROI cropping. Did this work for anyone else? @remekkinas How did this go for you?" } }