{ "id": 369364, "title": "RSNA Dataset Breakdown, Compiled list of External Datasets", "url": "/competitions/rsna-breast-cancer-detection/discussion/369364", "author_name": "gradientBoost", "post_date": "2022-11-30T00:40:01.506000", "votes": 23, "comment_count": 4, "views": 0, "content": "

Dataset Breakdown

\n

Total Patients:11913
\nTotal Unique Healthy Patients:11427
\nTotal Unique Cancer Patients:486
\nTotal Healthy Mammography Images: 53548
\nTotal Training images having Cancer: 1158

\n

Cancer Presence Breakdown (patients)

\n

Left Breast Only:242
\n Right Breast Only:238
\n Both Breasts:6

\n

Each patient has a total of 1-14 images in the training dataset

\n

*The provided training dataset is highly imbalanced! * Data Augmentation and external datasets can help train a decent classifier

\n

List of External Datasets

\n", "messages": [ { "id": 2049137, "postDate": "2022-11-30T00:40:01.507Z", "content": "

Dataset Breakdown

\n

Total Patients:11913
\nTotal Unique Healthy Patients:11427
\nTotal Unique Cancer Patients:486
\nTotal Healthy Mammography Images: 53548
\nTotal Training images having Cancer: 1158

\n

Cancer Presence Breakdown (patients)

\n

Left Breast Only:242
\n Right Breast Only:238
\n Both Breasts:6

\n

Each patient has a total of 1-14 images in the training dataset

\n

*The provided training dataset is highly imbalanced! * Data Augmentation and external datasets can help train a decent classifier

\n

List of External Datasets

\n", "rawMarkdown": "## Dataset Breakdown\nTotal Patients:11913 \nTotal Unique Healthy Patients:11427\nTotal Unique Cancer Patients:486 \nTotal Healthy Mammography Images: 53548\nTotal Training images having Cancer: 1158\n\n\n## Cancer Presence Breakdown (*patients*)\nLeft Breast Only:242\n Right Breast Only:238\n Both Breasts:6\n \nEach patient has a total of 1-14 images in the training dataset\n\n\n*The provided training dataset is **highly imbalanced**! * Data Augmentation and external datasets can help train a decent classifier\n\n## List of External Datasets\n- King Abdulaziz University Mammogram Dataset\nhttps://www.kaggle.com/datasets/asmaasaad/king-abdulaziz-university-mammogram-dataset\n\n- vindr.ai Dataset [5000 Images]\nhttps://physionet.org/content/vindr-mammo/1.0.0/\nhttps://vindr.ai/datasets/mammo\n\n- CBIS-DDSM Dataset\nhttps://www.kaggle.com/datasets/awsaf49/cbis-ddsm-breast-cancer-image-dataset \n\n- Mini MIAS Dataset\nhttp://peipa.essex.ac.uk/info/mias.html \n\n\n", "votes": 21 }, { "id": 2058646, "postDate": "2022-12-08T05:27:32.963Z", "content": "

not sure if this is useful
\nhttps://sites.duke.edu/mazurowski/resources/digital-breast-tomosynthesis-database/
\nWe are sharing a dataset of digital breast tomosynthesis (DBT) volumes for 5,060 patients. (1 TB)

\n

tomosynthesis = 3d Mammography
\n(i.e. you can create multiview)

\n

Training set (with truth):
\nThe training set consists of 19148 cases. This dataset will be representative of the technical properties (equipment, acquisition parameters, file format) and the nature of lesions in the validation and test sets. An associated Excel file in CSV format will include DBT scan identifier and the definition of the bounding box of all lesions.

\n
\n

tips:

\n

to test if external data is useful or not:

\n
    \n
  1. train a model using kaggle data
    2. \n
  2. validate using kaggle data, e.g. accuracy = ak
    3. \n
  3. validate using external data, e.g. accuracy = ae
    4. \n
\n

if ak is much better than ae, then the two dataset does not overlap. external data may not be useful.
\nbetter still, plot TSNE (or other distance embeddings to check distance of kaggle train, kaggle validation and external)

\n
\n

you can add external data incrementally, e.g. select pos(and/or neg) external data with prob score 0.3 to 0.6 first

\n

you can modify labels of external data if you think there are \"different\" from kaggle ground truth

", "rawMarkdown": "not sure if this is useful\nhttps://sites.duke.edu/mazurowski/resources/digital-breast-tomosynthesis-database/\nWe are sharing a dataset of digital breast tomosynthesis (DBT) volumes for 5,060 patients. (1 TB)\n\ntomosynthesis = 3d Mammography\n(i.e. you can create multiview)\n\nTraining set (with truth): \nThe training set consists of 19148 cases. This dataset will be representative of the technical properties (equipment, acquisition parameters, file format) and the nature of lesions in the validation and test sets. An associated Excel file in CSV format will include DBT scan identifier and the definition of the bounding box of all lesions. \n\n----\n\ntips:\n\nto test if external data is useful or not:\n1. train a model using kaggle data\n2. validate using kaggle data, e.g. accuracy = ak\n3. validate using external data, e.g. accuracy = ae\n\nif ak is much better than ae, then the two dataset does not overlap. external data may not be useful.\nbetter still, plot TSNE (or other distance embeddings to check distance of kaggle train, kaggle validation and external)\n\n---\n\nyou can add external data incrementally, e.g. select pos(and/or neg) external data with prob score 0.3 to 0.6 first\n\nyou can modify labels of external data if you think there are \"different\" from kaggle ground truth", "votes": 6 }, { "id": 2058593, "postDate": "2022-12-08T04:00:03.867Z", "content": "

Thank you for sharing! This is very informative.
\nI could find MIAS dataset on kaggle in here.

\n

I add some others from kaggle dataset.
\nINbreast
\nhttps://www.kaggle.com/datasets/tommyngx/inbreast2012

\n

CMMD
\nhttps://www.kaggle.com/datasets/tommyngx/cmmd2022

", "rawMarkdown": "Thank you for sharing! This is very informative.\nI could find MIAS dataset on kaggle in [here](https://www.kaggle.com/datasets/kmader/mias-mammography).\n\nI add some others from kaggle dataset.\nINbreast\nhttps://www.kaggle.com/datasets/tommyngx/inbreast2012\n\nCMMD\nhttps://www.kaggle.com/datasets/tommyngx/cmmd2022", "votes": 4 }, { "id": 2146047, "postDate": "2023-02-15T15:01:40.903Z", "content": "

I found the \"King Abdulaziz University Mammogram Dataset\" only used for research.
\nhttps://www.kaggle.com/datasets/asmaasaad/mammogram-dataset-kaumds?select=licence+agreement.docx

", "rawMarkdown": "I found the \"King Abdulaziz University Mammogram Dataset\" only used for research.\nhttps://www.kaggle.com/datasets/asmaasaad/mammogram-dataset-kaumds?select=licence+agreement.docx\n", "votes": 1 }, { "id": 2070189, "postDate": "2022-12-19T17:26:55.143Z", "content": "

There is a NCI Cancer Research Data Commons, which has lots of opensource medical imagery

\n

17 different databases are listed for breast body part

\n

https://portal.imaging.datacommons.cancer.gov/explore/

\n

@vbookshelf also created a dataset with yolov5 models 19 days ago, but if he mentioned it here, I didn't see it.

\n

The below-the-radar approach is quite intriguing.

\n

https://www.kaggle.com/datasets/vbookshelf/mammogram-mass-analyzer-v00

\n

based on the https://physionet.org/content/vindr-mammo/1.0.0/ mentioned above. Should probably highlight that one

", "rawMarkdown": "There is a NCI Cancer Research Data Commons, which has lots of opensource medical imagery\n\n17 different databases are listed for breast body part\n\nhttps://portal.imaging.datacommons.cancer.gov/explore/\n\n\n\n@vbookshelf also created a dataset with yolov5 models 19 days ago, but if he mentioned it here, I didn't see it.\n\nThe below-the-radar approach is quite intriguing.\n\nhttps://www.kaggle.com/datasets/vbookshelf/mammogram-mass-analyzer-v00\n\nbased on the https://physionet.org/content/vindr-mammo/1.0.0/ mentioned above. Should probably highlight that one", "votes": 1 } ], "comments": [ { "id": 2058646, "author_name": "hengck23", "author_url": "", "post_date": "2022-12-08T05:27:32.963000", "content": "

not sure if this is useful
\nhttps://sites.duke.edu/mazurowski/resources/digital-breast-tomosynthesis-database/
\nWe are sharing a dataset of digital breast tomosynthesis (DBT) volumes for 5,060 patients. (1 TB)

\n

tomosynthesis = 3d Mammography
\n(i.e. you can create multiview)

\n

Training set (with truth):
\nThe training set consists of 19148 cases. This dataset will be representative of the technical properties (equipment, acquisition parameters, file format) and the nature of lesions in the validation and test sets. An associated Excel file in CSV format will include DBT scan identifier and the definition of the bounding box of all lesions.

\n
\n

tips:

\n

to test if external data is useful or not:

\n
    \n
  1. train a model using kaggle data
    2. \n
  2. validate using kaggle data, e.g. accuracy = ak
    3. \n
  3. validate using external data, e.g. accuracy = ae
    4. \n
\n

if ak is much better than ae, then the two dataset does not overlap. external data may not be useful.
\nbetter still, plot TSNE (or other distance embeddings to check distance of kaggle train, kaggle validation and external)

\n
\n

you can add external data incrementally, e.g. select pos(and/or neg) external data with prob score 0.3 to 0.6 first

\n

you can modify labels of external data if you think there are \"different\" from kaggle ground truth

", "votes": 6, "replies": [] }, { "id": 2058593, "author_name": "tomoo inubushi", "author_url": "", "post_date": "2022-12-08T04:00:03.867000", "content": "

Thank you for sharing! This is very informative.
\nI could find MIAS dataset on kaggle in here.

\n

I add some others from kaggle dataset.
\nINbreast
\nhttps://www.kaggle.com/datasets/tommyngx/inbreast2012

\n

CMMD
\nhttps://www.kaggle.com/datasets/tommyngx/cmmd2022

", "votes": 4, "replies": [] }, { "id": 2146047, "author_name": "SenTran", "author_url": "", "post_date": "2023-02-15T15:01:40.903000", "content": "

I found the \"King Abdulaziz University Mammogram Dataset\" only used for research.
\nhttps://www.kaggle.com/datasets/asmaasaad/mammogram-dataset-kaumds?select=licence+agreement.docx

", "votes": 1, "replies": [] }, { "id": 2070189, "author_name": "@kaggleqrdl", "author_url": "", "post_date": "2022-12-19T17:26:55.143000", "content": "

There is a NCI Cancer Research Data Commons, which has lots of opensource medical imagery

\n

17 different databases are listed for breast body part

\n

https://portal.imaging.datacommons.cancer.gov/explore/

\n

@vbookshelf also created a dataset with yolov5 models 19 days ago, but if he mentioned it here, I didn't see it.

\n

The below-the-radar approach is quite intriguing.

\n

https://www.kaggle.com/datasets/vbookshelf/mammogram-mass-analyzer-v00

\n

based on the https://physionet.org/content/vindr-mammo/1.0.0/ mentioned above. Should probably highlight that one

", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "2049137": "## Dataset Breakdown\nTotal Patients:11913 \nTotal Unique Healthy Patients:11427\nTotal Unique Cancer Patients:486 \nTotal Healthy Mammography Images: 53548\nTotal Training images having Cancer: 1158\n\n\n## Cancer Presence Breakdown (*patients*)\nLeft Breast Only:242\n Right Breast Only:238\n Both Breasts:6\n \nEach patient has a total of 1-14 images in the training dataset\n\n\n*The provided training dataset is **highly imbalanced**! * Data Augmentation and external datasets can help train a decent classifier\n\n## List of External Datasets\n- King Abdulaziz University Mammogram Dataset\nhttps://www.kaggle.com/datasets/asmaasaad/king-abdulaziz-university-mammogram-dataset\n\n- vindr.ai Dataset [5000 Images]\nhttps://physionet.org/content/vindr-mammo/1.0.0/\nhttps://vindr.ai/datasets/mammo\n\n- CBIS-DDSM Dataset\nhttps://www.kaggle.com/datasets/awsaf49/cbis-ddsm-breast-cancer-image-dataset \n\n- Mini MIAS Dataset\nhttp://peipa.essex.ac.uk/info/mias.html \n\n\n", "2058646": "not sure if this is useful\nhttps://sites.duke.edu/mazurowski/resources/digital-breast-tomosynthesis-database/\nWe are sharing a dataset of digital breast tomosynthesis (DBT) volumes for 5,060 patients. (1 TB)\n\ntomosynthesis = 3d Mammography\n(i.e. you can create multiview)\n\nTraining set (with truth): \nThe training set consists of 19148 cases. This dataset will be representative of the technical properties (equipment, acquisition parameters, file format) and the nature of lesions in the validation and test sets. An associated Excel file in CSV format will include DBT scan identifier and the definition of the bounding box of all lesions. \n\n----\n\ntips:\n\nto test if external data is useful or not:\n1. train a model using kaggle data\n2. validate using kaggle data, e.g. accuracy = ak\n3. validate using external data, e.g. accuracy = ae\n\nif ak is much better than ae, then the two dataset does not overlap. external data may not be useful.\nbetter still, plot TSNE (or other distance embeddings to check distance of kaggle train, kaggle validation and external)\n\n---\n\nyou can add external data incrementally, e.g. select pos(and/or neg) external data with prob score 0.3 to 0.6 first\n\nyou can modify labels of external data if you think there are \"different\" from kaggle ground truth", "2058593": "Thank you for sharing! This is very informative.\nI could find MIAS dataset on kaggle in [here](https://www.kaggle.com/datasets/kmader/mias-mammography).\n\nI add some others from kaggle dataset.\nINbreast\nhttps://www.kaggle.com/datasets/tommyngx/inbreast2012\n\nCMMD\nhttps://www.kaggle.com/datasets/tommyngx/cmmd2022", "2146047": "I found the \"King Abdulaziz University Mammogram Dataset\" only used for research.\nhttps://www.kaggle.com/datasets/asmaasaad/mammogram-dataset-kaumds?select=licence+agreement.docx\n", "2070189": "There is a NCI Cancer Research Data Commons, which has lots of opensource medical imagery\n\n17 different databases are listed for breast body part\n\nhttps://portal.imaging.datacommons.cancer.gov/explore/\n\n\n\n@vbookshelf also created a dataset with yolov5 models 19 days ago, but if he mentioned it here, I didn't see it.\n\nThe below-the-radar approach is quite intriguing.\n\nhttps://www.kaggle.com/datasets/vbookshelf/mammogram-mass-analyzer-v00\n\nbased on the https://physionet.org/content/vindr-mammo/1.0.0/ mentioned above. Should probably highlight that one" } }