\nIn fact, mammography is the imaging modality with the most resolution, requiring specially crafted (and FDA approved) monitors for diagnostic evaluation :
\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F8926747%2F29305461307f37a6e9b555419c3c686b%2Ffoto_no_exif.jpg?generation=1672055892771870&alt=media\")
\n(Table from [4])
Hello everyone 👋
\nI’m João Abrantes, a Radiologist from Portugal.
\nIn my clinical setting I mainly work with breast imaging, and mammography analysis consists of a large part of my day-to-day activity.
\nA lot of great topics have been discussed in this competition, and I wanted to offer some insight into the mammography analysis and some tips and tricks that can (hopefully) be helpful for the development of better algorithms.
\nThis is part 1 of a series of topics that I'll be writing in the next few days.
👉I recommend that everyone read the topic by Ian Pan (https://www.kaggle.com/competitions/rsna-breast-cancer-detection/discussion/369262) for some very helpful intro related to Radiology, Breast Cancer and the Screening Workflow.
\n✌️Check part 2 of the guide here: https://www.kaggle.com/competitions/rsna-breast-cancer-detection/discussion/374946
\nMAMMOGRAPHY
\nMammography is one of the most technically challenging areas of radiography, requiring high spatial resolution, excellent soft-tissue contrast, and low radiation dose.
\nScreening programs (“Screening is a way of identifying apparently healthy people who may have an increased risk of a particular condition” [1]). In a Canadian study, breast cancer screening lead to a decrease on breast cancer mortality - reduced by 44% among screened women aged 40 to 49, 40% in screened women aged 50 to 59, 42% in screened women aged 60 to 69, and 35% in screened women aged 70 to 79 compared with unscreened women [2].
Digital mammography requires a very high spatial resolution, in order to capture the features of breast lesions (contours, density, spiculation, distortions of the normal breast architecture) and for microcalcifications (a type of finding that can frequently represent a subset of breast cancer-DCIS).
\n🔥My take: As such, for this specific purpose, it is expected that higher resolution images could contribute for a better algorithm performance ( as some topics already covered, based on the processed dicom data at different resolutions, from the thread : https://www.kaggle.com/datasets/radek1/rsna-mammography-images-as-pngs
\nPRE-PROCESSING
\nSeveral techniques for image data augmentation are well established and part of popular libraries, and the severe imbalance in this dataset (healthy vs. cancer cases) forces us to tackle this problem.
\nThe topic (https://www.kaggle.com/competitions/rsna-breast-cancer-detection/discussion/372567) lists some very helpful data augmentation techniques that could be of use in this setting.
\nThe majority of breast carcinomas have the mammographic appearance of a radiating structure, either a definite stellate/spiculated lesion or architectural distortion with no central tumor mass. The most typical mammographic appearance of breast carcinoma is a stellate lesion, i.e.a solid central tumor mass surrounded by a radiating structure. [5]
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F8926747%2F8e12f0b28b4703d92d509e0e3abe6b03%2Ffoto_no_exif%20(1).jpg?generation=1672056191633107&alt=media\")
\nImage by [5]
🔥My take: The use of augmentations techniques that alter the image dimension ratio can lead to the loss of these characteristic morphologic signs, and could utimattly lead to a lower algorithm performance. The pre-processing of image data can be one of the most important aspects for a good result in this challenge.
\nThank you for reading so far.
\nThis is my first input to this discussion and in the next few days I’ll be trying to contribute and give some more insight into the clinical aspect of this challenge.
\nSome other topics to cover:
👇Feel free to ask away in the comments, I'll do my best to try and positively contribute for the discussion!👇
\nReferences
\n[1] NHS, \"NHS Screening,\" 20 July 2021. [Online]. Available: https://www.nhs.uk/conditions/nhs-screening/.
\n[2] P. N. W. L. e. Coldman A, \"Breast cancer mortality after screening mammography in British Columbia women,\" Int J Cancer , vol. 120, p. 1076–1080, 2007..
\n[3] J. A. Seibert, \"Society for Imaging Informatics in Medicine,\" Society for Imaging Informatics in Medicine, 2021. [Online]. Available: https://siim.org/page/archiving_chapter2.
\n[4] H. O. Ken Compton, \"OTech Inc,\" OTech Inc, [Online]. Available: https://otechimg.com/publications/pdf/wp_medical_image_monitors.pdf. [Accessed 2022].
\n[5] P. B. D. Laszlo Tabar, Teaching Atlas of Mammography, Thieme Publishing Group, 2012.
Hello everyone 👋
\nI’m João Abrantes, a Radiologist from Portugal.
\nIn my clinical setting I mainly work with breast imaging, and mammography analysis consists of a large part of my day-to-day activity.
\nA lot of great topics have been discussed in this competition, and I wanted to offer some insight into the mammography analysis and some tips and tricks that can (hopefully) be helpful for the development of better algorithms.
\nThis is part 1 of a series of topics that I'll be writing in the next few days.
👉I recommend that everyone read the topic by Ian Pan (https://www.kaggle.com/competitions/rsna-breast-cancer-detection/discussion/369262) for some very helpful intro related to Radiology, Breast Cancer and the Screening Workflow.
\n✌️Check part 2 of the guide here: https://www.kaggle.com/competitions/rsna-breast-cancer-detection/discussion/374946
\nMAMMOGRAPHY
\nMammography is one of the most technically challenging areas of radiography, requiring high spatial resolution, excellent soft-tissue contrast, and low radiation dose.
\nScreening programs (“Screening is a way of identifying apparently healthy people who may have an increased risk of a particular condition” [1]). In a Canadian study, breast cancer screening lead to a decrease on breast cancer mortality - reduced by 44% among screened women aged 40 to 49, 40% in screened women aged 50 to 59, 42% in screened women aged 60 to 69, and 35% in screened women aged 70 to 79 compared with unscreened women [2].
Digital mammography requires a very high spatial resolution, in order to capture the features of breast lesions (contours, density, spiculation, distortions of the normal breast architecture) and for microcalcifications (a type of finding that can frequently represent a subset of breast cancer-DCIS).
\n🔥My take: As such, for this specific purpose, it is expected that higher resolution images could contribute for a better algorithm performance ( as some topics already covered, based on the processed dicom data at different resolutions, from the thread : https://www.kaggle.com/datasets/radek1/rsna-mammography-images-as-pngs
\nPRE-PROCESSING
\nSeveral techniques for image data augmentation are well established and part of popular libraries, and the severe imbalance in this dataset (healthy vs. cancer cases) forces us to tackle this problem.
\nThe topic (https://www.kaggle.com/competitions/rsna-breast-cancer-detection/discussion/372567) lists some very helpful data augmentation techniques that could be of use in this setting.
\nThe majority of breast carcinomas have the mammographic appearance of a radiating structure, either a definite stellate/spiculated lesion or architectural distortion with no central tumor mass. The most typical mammographic appearance of breast carcinoma is a stellate lesion, i.e.a solid central tumor mass surrounded by a radiating structure. [5]
\nImage by [5]
🔥My take: The use of augmentations techniques that alter the image dimension ratio can lead to the loss of these characteristic morphologic signs, and could utimattly lead to a lower algorithm performance. The pre-processing of image data can be one of the most important aspects for a good result in this challenge.
\nThank you for reading so far.
\nThis is my first input to this discussion and in the next few days I’ll be trying to contribute and give some more insight into the clinical aspect of this challenge.
\nSome other topics to cover:
👇Feel free to ask away in the comments, I'll do my best to try and positively contribute for the discussion!👇
\nReferences
\n[1] NHS, \"NHS Screening,\" 20 July 2021. [Online]. Available: https://www.nhs.uk/conditions/nhs-screening/.
\n[2] P. N. W. L. e. Coldman A, \"Breast cancer mortality after screening mammography in British Columbia women,\" Int J Cancer , vol. 120, p. 1076–1080, 2007..
\n[3] J. A. Seibert, \"Society for Imaging Informatics in Medicine,\" Society for Imaging Informatics in Medicine, 2021. [Online]. Available: https://siim.org/page/archiving_chapter2.
\n[4] H. O. Ken Compton, \"OTech Inc,\" OTech Inc, [Online]. Available: https://otechimg.com/publications/pdf/wp_medical_image_monitors.pdf. [Accessed 2022].
\n[5] P. B. D. Laszlo Tabar, Teaching Atlas of Mammography, Thieme Publishing Group, 2012.
In the vindr dataset, it categories as follows:
\nhttps://physionet.org/content/vindr-mammo/1.0.0/finding_annotations.csv
\nMass
\n['Global Asymmetry']
\n['Architectural Distortion']
\n'Nipple Retraction
\nSuspicious Calcification'
\nFocal Asymmetry'
\n['Asymmetry']
\n['Skin Thickening']
\n['Suspicious Lymph Node']
Do these categories make sense to you? It'd be interesting to get a good radiologist guide on how to spot each of these. Do you know if one exists?
", "rawMarkdown": "@jabrantes \n\nIn the vindr dataset, it categories as follows:\n\nhttps://physionet.org/content/vindr-mammo/1.0.0/finding_annotations.csv\n\nMass\n['Global Asymmetry']\n['Architectural Distortion']\n'Nipple Retraction\nSuspicious Calcification'\nFocal Asymmetry'\n['Asymmetry']\n['Skin Thickening']\n['Suspicious Lymph Node']\n\nDo these categories make sense to you? It'd be interesting to get a good radiologist guide on how to spot each of these. Do you know if one exists?\n\n\n", "votes": 2, "replies": [ { "id": 2076604, "postDate": "2022-12-26T16:18:30.930Z", "content": "Hello @kaggleqrdl
\nThose categories are associated findings of masses worrisome for cancer, and can aid in the diagnosis of a malignant lesions when present.
\nTaking into account the topic of breast asymmetry:
\nFrom the Americal College of Radiology BI-RADS atlas, the definitions of breast asymmetry are as follows:
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F8926747%2F43bad73d1f8fe05ef76ca74c0d42d42f%2Fasymmetry.png?generation=1672070751508391&alt=media\")
\nSome examples (source: Breast Imaging: The Requisites (2016) - Debra Ikeda, Kanae Kawai Miyake):
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F8926747%2F9e9888b610df818109b73e4bcb10bb8c%2Fasymmetry%20exemples.png?generation=1672070930751074&alt=media\")
So, to reach this definition one needs to compare both the left and right breast or different time points of the same breast (in the case of a developing asymmetry).
\n👉Maybe this can be of use for the current competition: Can the comparison between both sides increase the performance of the algorithm?👈
Thanks, this was quite useful.
", "rawMarkdown": "Thanks, this was quite useful. ", "votes": 1 } ] }, { "id": 2076612, "postDate": "2022-12-26T16:24:31.727Z", "content": "these are standard reporting in mammography screening.
\nyou can google or ask chatGPT about it
https://radiologyassistant.nl/breast/bi-rads/bi-rads-for-mammography-and-ultrasound-2013
\nhttps://www.facebook.com/tahonnour/photos/guide-for-breast-calcifications/1164548206914164
In the vindr dataset, it categories as follows:
\nhttps://physionet.org/content/vindr-mammo/1.0.0/finding_annotations.csv
\nMass
\n['Global Asymmetry']
\n['Architectural Distortion']
\n'Nipple Retraction
\nSuspicious Calcification'
\nFocal Asymmetry'
\n['Asymmetry']
\n['Skin Thickening']
\n['Suspicious Lymph Node']
Do these categories make sense to you? It'd be interesting to get a good radiologist guide on how to spot each of these. Do you know if one exists?
", "votes": 2, "replies": [ { "id": 2076604, "author_name": "JAbrantes", "author_url": "", "post_date": "2022-12-26T16:18:30.930000", "content": "Hello @kaggleqrdl
\nThose categories are associated findings of masses worrisome for cancer, and can aid in the diagnosis of a malignant lesions when present.
\nTaking into account the topic of breast asymmetry:
\nFrom the Americal College of Radiology BI-RADS atlas, the definitions of breast asymmetry are as follows:
\n
\nSome examples (source: Breast Imaging: The Requisites (2016) - Debra Ikeda, Kanae Kawai Miyake):
\n
So, to reach this definition one needs to compare both the left and right breast or different time points of the same breast (in the case of a developing asymmetry).
\n👉Maybe this can be of use for the current competition: Can the comparison between both sides increase the performance of the algorithm?👈
Thanks, this was quite useful.
", "votes": 1, "replies": [] } ] }, { "id": 2076612, "author_name": "hengck23", "author_url": "", "post_date": "2022-12-26T16:24:31.727000", "content": "these are standard reporting in mammography screening.
\nyou can google or ask chatGPT about it
https://radiologyassistant.nl/breast/bi-rads/bi-rads-for-mammography-and-ultrasound-2013
\nhttps://www.facebook.com/tahonnour/photos/guide-for-breast-calcifications/1164548206914164