You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Terms and Conditions for Using the Dataset
1. Acceptance of Terms
Accessing and using the dataset implies your agreement to the terms and conditions. If you disagree with any part, please refrain from using the dataset.
Refer to the respective licenses of the original datasets for permitted usage and terms.
2. Attribution
Cite AADD as well as the original sources of the data, and acknowledge the providers in any publications resulting from its use.
3. Disclaimer
The dataset is provided "as is" without warranty of any kind, either expressed or implied, including but not limited to the accuracy or completeness of the data.
4. Limitation of Liability
Under no circumstances will the dataset providers be liable for any claims or damages resulting from your use of the dataset.
5. Amendments
The terms and conditions may be updated at any time; continued use of the dataset signifies acceptance of the new terms.
Consent:
Accessing and using the dataset signifies your acknowledgment and agreement to these terms and conditions.
Log in or Sign Up to review the conditions and access this dataset content.
Advancing Anomaly Detection Dataset
This dataset is designed to support the evaluation of medical anomaly detection models. It contains brain MRI and abdominal CT volumetric data, which come from the training data of the medical out-of-distribution (MOOD) analysis challenge. In contrast to most other datasets, we use synthetic alterations to create both normal and abnormal classes. By introducing synthetic elements into both normal and anomalous classes, we can reduce the domain shift that sometimes occurs when only anomalous samples are altered. Ideally, this gives us greater control over which aspects the model must detect as anomalous. Using synthetic anomalies also grants us (nearly) complete labels at no additional cost. We hope this dataset can be a useful resource for scrutinizing anomaly detection methods with more precise specification of anomalies and labels, while minimizing data collection and annotation costs.
This dataset is released in tandem with a call for papers, advancing anomaly detection in radiology, in Frontiers in Radiology. Authors submitting an article proposing a new anomaly detection method are required to evaluate their method on at least one modality within this dataset, preferably both. If this dataset is of interest to you, please consider submitting an article.
Brain
Below are examples from the brain dataset. Normal samples have a band of spatial frequencies removed and are
rotated about a random axis by an angle in the range of [30,60] degrees.
Anomalous samples are over-rotated (by an angle in the range of [75,105] degrees); under-rotated (by an angle
in the range of [0,15] degrees); or rotated normally but including additional spatial frequencies; or with
spatial frequencies added only within a certain region. Labels below anomalous samples indicate the region of the anomaly.
Abdominal
Below are examples from the abdominal dataset. Normal samples include a new synthetic organ, the segmentation
mask below indicates its location (not abnormality).
Anomalous samples include added texture to the synthetic organ; split
morphology into two smaller organs; normal organ in an abnormal location; and the complete absence
of the organ. The labels below the anomalous samples indicate the region of the anomaly.
Usage and Evaluation Specifics
Dataset Structure
Below is an overview of the brain and abdom directories. Both contain a train directory as well as a test0 directory, containing normal test samples, and test1-4 directories, containing the four types of anomalies depicted in the respective figures above.
AADD
│
└───brain
│ │
│ └───train
│ │ │ aadd_0*.nii.gz
│ │ │ ...
│ │
│ └───test
│ │
│ └───test0
│ │ │ aadd_0*.nii.gz
│ │ │ ...
│ │
│ └───test1
│ │ │ aadd_0*.nii.gz
│ │ │ aadd_label_0*.nii.gz
│ │ │ ...
│ │
│ └───test2
│ │ │ aadd_0*.nii.gz
│ │ │ aadd_label_0*.nii.gz
│ │ │ ...
│ │
│ └───test3
│ │ │ aadd_0*.nii.gz
│ │ │ aadd_label_0*.nii.gz
│ │ │ ...
│ │
│ └───test4
│ │ aadd_0*.nii.gz
│ │ aadd_label_0*.nii.gz
│ │ ...
│
│
└───abdom
│
└───train
│ │ aadd_0*.nii.gz
│ │ ...
│
└───test
│
└───test0
│ │ aadd_0*.nii.gz
│ │ ...
│
└───test1
│ │ aadd_0*.nii.gz
│ │ aadd_label_0*.nii.gz
│ │ ...
│
└───test2
│ │ aadd_0*.nii.gz
│ │ aadd_label_0*.nii.gz
│ │ ...
│
└───test3
│ │ aadd_0*.nii.gz
│ │ aadd_ignore_0*.nii.gz
│ │ aadd_label_0*.nii.gz
│ │ ...
│
└───test4
│ aadd_0*.nii.gz
│ aadd_label_0*.nii.gz
│ ...
Training
For training, we request all users to use only the data within the train directories. Data within the test directories should not be used for training or validation. This helps to keep all reported scores comparable across papers. This is particularly important for authors intending to submit to the Frontiers call for papers. If you wish to use test data for validation please specify the details clearly in any tables reporting your performance.
Testing/Evaluation and Annotation Notes
For testing, all data in test0 directories are normal test samples. No annotation label is provided for these samples but you may
assume the label is zero everywhere. The data in test1-4 contains anomalous test samples and their associated labels.
***Important***
Some labels are not binarized so that you can examine the differences more clearly; as such, please use the following (or equivalent) before computing your scores:
label = label!=0
Note that abdom/test/test3 contains an additional mask (aadd_ignore_0*.nii.gz). For this case, your predictions can be multiplied as follows (or equivalent) before computing your scores:
pred_00007 = pred_00007*np.logical_not(aadd_ignore_00007)
As for metrics, we ask that all users report at least AUROC and AP at both sample and pixel level, if their method permits. We recommend that users report total performance (test0 vs test1-4) as well as performance on individual types of anomalies, by randomly sampling from test0 to get balanced classes (e.g. test0 vs test1, 30 samples each).
Limitations and Feedback
This dataset is obviously not without its limitations and we would appreciate your input to help us improve future iterations.
If you would like to provide feedback, consider filling out our community input form.
Citation
TBD - a report with more details will be uploaded shortly
Original Sources of the Data
The original data comes from the MOOD challenge training data and its respective sources:
1. @article{zimmerer2022mood,
title={MOOD 2020: A public Benchmark for Out-of-Distribution Detection and Localization on medical Images},
author={Zimmerer, David and Full, Peter M and Isensee, Fabian and J{\"a}ger, Paul and Adler, Tim and Petersen, Jens and K{\"o}hler, Gregor and Ross, Tobias and Reinke, Annika and Kascenas, Antanas and others},
journal={IEEE Transactions on Medical Imaging},
volume={41},
number={10},
pages={2728--2738},
year={2022},
publisher={IEEE}
}
2. @article{smith2015data,
title={Data from CT\_COLONOGRAPHY},
author={Smith, K and Clark, K and Bennett, W and Nolan, T and Kirby, J and Wolfsberger, M and Moulton, J and Vendt, B and Freymann, J},
journal={Cancer Imaging Arch},
volume={10},
pages={K9},
year={2015}
}
3. @article{van2013wu,
title={The WU-Minn human connectome project: an overview},
author={Van Essen, David C and Smith, Stephen M and Barch, Deanna M and Behrens, Timothy EJ and Yacoub, Essa and Ugurbil, Kamil and Wu-Minn HCP Consortium and others},
journal={Neuroimage},
volume={80},
pages={62--79},
year={2013},
publisher={Elsevier}
}
- Downloads last month
- 1