This site uses cookies to enhance your browser experience. By clicking "Accept", you consent to your use of cookies. Cookie Policy
ACR AI-LAB™ To accelerate the development and adoption of artificial intelligence (AI) in clinical practice, radiologists need to be empowered to create AI tools at their own institutions, to meet their own patient needs. ACR AI-LAB™ offers radiologists tools designed to help them learn the basics of AI and participate directly in the creation, validation and use of healthcare AI.
1891 Preston White Dr.
Reston, VA 20191
The ACR Data Science Institute has developed the ACR AI-LAB™, a data science toolkit designed to democratize AI by empowering radiologists to develop algorithms at their own institutions, using their own patient data, to meet their own clinical needs.
Welcome to the ACR AI-LAB™ Define. Browse defined AI use cases, comment on published use cases, and submit a new use case for review.
1.Browse defined AI use cases
2.Search, filter, or sort through all of the use cases
3.Click on the use case name under the USE CASE column to comment on a published use case.
4.Click on REQUEST TO DRAFT A USE CASE under the REQUEST column to request to draft a use case that is not yet published.
5.Click on SUBMIT A NEW USE CASE and fill out the form to submit a new AI use case for the ACR to review.
Welcome to the ACR AI-LAB™ Annotate. Within each use case, you can annotate multiple scans to develop a data set.
1. Select a use case.
2. Select a data set to annotate.
3. Use the image viewing tools to examine the scan.
4. Classify each image by selecting one of the categories.
5. Click NEXT to save your annotation and view the next scan.
6. Click FINISH to finish annotating and return to the home page.
Welcome to the ACR AI-LAB™ My Annotations. Annotate at least 100 of the 5,000 studies to use to build your own breast density classification model.
1. Use the image viewing tools to examine the scan.
2. Classify each image by selecting one of the categories.
3. Click SAVE AND CONTINUE to save your annotation and view the next scan.
4. Click SKIP to skip a study. You can always go back to review your skipped studies by clicking on the first dropdown in the upper left corner.
5. View your annotated studies by clicking on the second dropdown in the upper left corner. Change your annotation by selecting a new classification and clicking SAVE AND CONTINUE.
6. Click FINISH to finish annotating and return to the leaderboard. You can return at any time to continue annotating.
If you experience any issues during the challenge, please let us know by emailing ailabsupport@acr.org.
Welcome to the ACR AI-LAB™ Run. You can run a model and see the prediction by using a prepopulated image or by uploading your own image.
Step 1: Define-AI generates CDEs for a specific use case.
Step 2: An Assist Module is created for each Define-AI use case which encodes CDEs and associated Report Text as XML
Step 3: The Assist module can be represented in any UI such as:
Step 4: When the algorithm generates output, the associated Report Text is also generated. The text can be customized to local preference (e.g., can be blank if result is negative).
Welcome to the ACR AI-LAB™ Publish. You can publish a model you designed on the Create page to a destination like an AI marketplace.
Welcome to the ACR AI-LAB™ Assess. You can assess the performances of all deployed models.
Statistic that measures the agreement between two raters that classify items into mutually exclusive categories and takes into account the probability of random agreement. If two raters are in complete agreement, then the kappa score could be close to 1. If two raters are in complete disagreement, then the kappa score could be close to 0 or even negative if they are agreeing less than the probability of random agreement.
A table that is used to display the performance of a classification model on a test data set as compared to the ground truth. Each square has a unique row and column combination. The square in row 1 and column 2 says that the ground truth classified those test data points as category 1 but the classification model classified those same test data points as category 2. All of the squares along the diagonal of the table display the number of test cases when the ground truth and classification model are in agreement. All squares off the diagonal represent any disagreement between the ground truth and classification model; the further away from the diagonal, the larger the disagreement. If there is complete agreement between the ground truth and classification model, then all of the off diagonal squares will be zero. If there is complete disagreement between the ground truth and classification model, then all of the diagonal squares will be zero.
Change the minimum probability threshold value required to classify the study as the positive class. If the probability that a study is in the positive class is greater than or equal to the threshold value, then the study is classified as positive. A threshold of 0.75 requires that the model be 75% sure that the study is in the positive class in order for it to be classified as positive. Whereas a threshold of 0.1 requires that the model only be 10% sure that the study is in the positive class in order for it to be classified as positive. Moving the threshold slider changes the statistics calculated for the specific data element.
A table that is used to display the performance of a classification model on a test data set as compared to the ground truth. Each square has a unique row and column combination. The square in row 1 and column 2 says that the ground truth classified those test data points as category 2 but the classification model classified those same test data points as category 1. All of the squares along the diagonal of the table display the number of test cases when the ground truth and classification model are in agreement. All squares off the diagonal represent any disagreement between the ground truth and classification model; the further away from the diagonal, the larger the disagreement. If there is complete agreement between the ground truth and classification model, then all of the off diagonal squares will be zero. If there is complete disagreement between the ground truth and classification model, then all of the diagonal squares will be zero.
Statistic that measures the agreement between two raters that classify items into mutually exclusive categories and takes into account the probability of random agreement. If two raters are in complete agreement, then the kappa score could be close to 1. If two raters are in complete disagreement, then the kappa score could be close to 0 or even negative if they are agreeing less than the probability of random agreement.
(Area Under the the Receiver Operating Characteristic (ROC) Curve or AUC, also known as the C-statistic) Performance measurement for classification problems that measures how well a model separates categories. An AUC score close to 1 means that the model is good at separating categories; it classifies 0s as 0s and 1s as 1s. An AUC score close to 0 means that the model is separating into opposite categories; it classifies 0s as 1s and 1s as 0s. An AUC score close to 0.5 means that the model is no better at separating categories than random chance.
Statistic that measures the ability of a binary classification model to correctly identify the positive (present) cases. It is the ratio of true positives (both the radiologist and model classify the case as positive) to ground truth positives (all cases the radiologist classifies as positive). Also commonly referred to as true positive rate. A model with 100% sensitivity identifies all positive cases and has few false negatives. However, sensitivity does not factor in false positives.
Statistic that measures the ability of a binary classification model to correctly identify the negative (absent) cases. It is the ratio of true negatives (both the radiologist and model classify the case as negative) to ground truth negatives (all cases the radiologist classifies as negative). Also commonly referred to as selectivity or true negative rate. A model with 100% specificity identifies all negative cases and has few false positives. However, specificity does not factor in false negatives.
Statistic that measures the agreement between two raters that classify items into mutually exclusive categories and takes into account the probability of random agreement. If two raters are in complete agreement, then the kappa score could be close to 1. If two raters are in complete disagreement, then the kappa score could be close to 0 or even negative if they are agreeing less than the probability of random agreement.
A scatterplot that compares the size of the predicted volumes to the size of the ground truth volumes. If all of the points fall on the dotted line, then the size of the predicted volumes is equal to the size of the ground truth volumes. If the points fall below the dotted line, then the model is generally predicting smaller volumes than the ground truth. If the points fall above the dotted line, then the model is generally predicting larger volumes.
The Bland-Altman plot shows the agreement between two quantitative measurements. The x-axis shows the average of the two measures and the y-axis shows the difference between the two measures. The mean difference is shown as a solid horizontal line and indicates the bias between the two measurements. If mean difference line is at 0, then on average, there is no bias between the measurements. Additionally, there are two more solid horizontal lines indicating the ±2 standard deviations of the mean difference (about 95% of the points should fall between the two lines). The Bland-Altman plot is helpful in discovering biases in quantitative measurements.
Statistic that measures the average similarity between two sets (segmentations or bounding boxes for example). A dice coefficient close to 1 signifies the sets are very similar; there is a large overlap between the predicted set and the ground truth set. A dice coefficient close to 0 signifies the sets are very dissimilar; there is very little overlap. The dice coefficient is positively correlated with Intersection over Union statistic. Generally, the mean dice coefficient will be higher for larger ground truth volumes (the radiologist's segmentation is large compared to the original image) than for smaller ground truth volumes (the radiologist's segmentation is small).
Statistic that measures the Average Precision of the volumes at different Intersection over Union thresholds. Average Precision is the ratio of true positives to true positives, false positives, and false negatives. Average Precision is not concerned with true negatives. A Mean Average Precision close to 1 indicates that on average, there is a large overlap of predicted volumes and ground truth volumes. A Mean Average Precision close to 0 indicates that the predicted volumes are generally missing the ground truth volumes. Because this statistic is for volumes and not classes, another statistic (Intersection over Union) is used to calculate which volumes are labeled true positives. An IoU Threshold (T) is used as the minimum IoU required to label a volume as a true positive. For example, if the IoU Threshold is set to T=0.5, then a predicted volume that overlaps 75% with the ground truth (IoU = 0.75 which is greater than T) is considered a true positive. However, a predicted volume that overlaps 49% with the ground truth (IoU = 0.49 which is less than T) is considered a false negative. The Mean Average Precision statistic is the mean Average Precision of all images over the range of IoU Thresholds.
Statistic that measures the average ratio of the intersection (volume of overlap) and the union (combined volumes) of the predicted and ground truth volumes. If the prediction matches the ground truth exactly, then the overlap and the union would be the same, resulting in an Intersection over Union (IoU)=1. If the prediction completely misses the ground truth, then the overlap is zero, resulting in an Intersection over Union (IoU)=0. The Mean IoU is the average intersection over union of all the bounding boxes.
Sample the studies randomly, in the distribution they are found in the original dataset.
Sample only the positive cases in the training dataset.
Sample the studies so that each class is represented equally – 25% almost entirely fat, 25% scattered fibroglandular densities etc.
Sigmoid Focal Loss reduces the importance of easy cases and increases the importance of the more difficult cases. This is very helpful when using imbalanced datasets - when there are a lot more of one class than the other. Sigmoid Focal Loss allows the model to focus on the more difficult cases.
Predicted Probability over Time uses the Study Date field in the DICOM header to monitor the model's predictions over time. The y-axis is the probability that the case is in the positive class (present) according to the model. The horizontal line indicates the Binary Threshold; the model classifies all cases above the threshold line as positive (present) and all cases below as negative (absent). In instances where the radiologist disagrees with the model's classification, the color will indicate if it is a false positive or a false negative.
Receiver operating Characteristic (RoC) Curve shows the trade-off between sensitivity and specificity for all thresholds. The blue circle will move depending on the Binary Threshold. A curve that is close to diagonal indicates that the model is no better than random chance. A curve that hugs the top left corner indicates good model performance. The area under the RoC curve is the AUC score.
Welcome to the ACR AI-LAB™ Evaluate. Within each use case, you can test multiple models on the same test data set and compare the performances.
Statistic that measures the agreement between two raters that classify items into mutually exclusive categories and takes into account the probability of random agreement. If two raters are in complete agreement, then the kappa score could be close to 1. If two raters are in complete disagreement, then the kappa score could be close to 0 or even negative if they are agreeing less than the probability of random agreement.
(Area under the curve of the Receiver operating characteristic) Performance measurement for classification problems that measures how well a model separates categories. An AUC score close to 1 means that the model is good at separating categories; it classifies 0s as 0s and 1s as 1s. An AUC score close to 0 means that the model is separating into opposite categories; it classifies 0s as 1s and 1s as 0s. An AUC score close to 0.5 means that the model is no better at separating categories than random chance.
A table that is used to display the performance of a classification model on a test data set as compared to the ground truth. Each square has a unique row and column combination. The square in row 1 and column 2 says that the ground truth classified those test data points as category 1 but the classification model classified those same test data points as category 2. All of the squares along the diagonal of the table display the number of test cases when the ground truth and classification model are in agreement. All squares off the diagonal represent any disagreement between the ground truth and classification model; the further away from the diagonal, the larger the disagreement. If there is complete agreement between the ground truth and classification model, then all of the off diagonal squares will be zero. If there is complete disagreement between the ground truth and classification model, then all of the diagonal squares will be zero.
Welcome to the ACR AI-LAB™ Create. Within each use case, you can design your own AI models and train, test, and run them on data sets.
The size of the training dataset can affect the accuracy of the model. In general, the model will be more accurate as the training dataset size increases.
The dataset that contains the prepared images that you train the model with.
If no data sets are listed, you will need to prepare them first.
Apply additional transformations, such as rotations, flips, or stretching, to images input to the network.
Structure of the neural network.
Technique for selecting the studies that the model sees at each training step.
How the model knows what to learn
Whether or not the model has been trained on a different task before training on a new task
The number of training examples the model learns from during a single step. If the batch size is too small, it can lead to chaotic training because the model doesn't see enough examples at once. If the batch size is too large, the training may fail due to memory constraints of the server. If the training keeps failing, try the smallest batch size possible and increase until the memory limit is exceeded.
How much the model should change while learning from the training data. A learning rate that is too small can lead to long training times and could cause the model to get 'stuck'. A learning rate that is too high can lead to chaotic training because the model is changing too much.
Number of epochs the model will wait for improvement before early stopping. If early stopping is True, then the model will continue to train for a minimum improvement interval. If no improvement is made, then the model will stop training early. If there is improvement, the interval restarts. If early stopping is False, the model will not stop training early.
If there is early stopping, the model will stop learning once its performance is no longer improving.
The metric being used to measure the performance of the algorithm.
Statistic that measures the agreement between two raters that classify items into mutually exclusive categories and takes into account the probability of random agreement. If two raters are in complete agreement, then the kappa score could be close to 1. If two raters are in complete disagreement, then the kappa score could be close to 0 or even negative if they are agreeing less than the probability of random agreement.
(Area under the curve of the Receiver operating characteristic) Performance measurement for classification problems that measures how well a model separates categories. An AUC score close to 1 means that the model is good at separating categories; it classifies 0s as 0s and 1s as 1s. An AUC score close to 0 means that the model is separating into opposite categories; it classifies 0s as 1s and 1s as 0s. An AUC score close to 0.5 means that the model is no better at separating categories than random chance.
A table that is used to display the performance of a classification model on a test data set as compared to the ground truth. Each square has a unique row and column combination. The square in row 1 and column 2 says that the ground truth classified those test data points as category 1 but the classification model classified those same test data points as category 2. All of the squares along the diagonal of the table display the number of test cases when the ground truth and classification model are in agreement. All squares off the diagonal represent any disagreement between the ground truth and classification model; the further away from the diagonal, the larger the disagreement. If there is complete agreement between the ground truth and classification model, then all of the off diagonal squares will be zero. If there is complete disagreement between the ground truth and classification model, then all of the diagonal squares will be zero.
Welcome to the ACR AI-LAB™ Train My Models. Design and train models using the data annotated on My Annotations. Submit models to the leaderboard.
The size of the training dataset can affect the accuracy of the model. In general, the model will be more accurate as the training dataset size increases.
Apply additional transformations, such as rotations, flips, or stretching, to images input to the network.
.......
Structure of the neural network.
The ratio of each class that the model sees at each training step
How the model knows what to learn
Whether or not the model has been trained on a different task before training on a new task
If there is early stopping, the model will stop learning once its performance is no longer improving.
The number of times the model learns from the training dataset.
Statistic that measures the agreement between two raters that classify items into mutually exclusive categories and takes into account the probability of random agreement. If two raters are in complete agreement, then the kappa score could be close to 1. If two raters are in complete disagreement, then the kappa score could be close to 0 or even negative if they are agreeing less than the probability of random agreement.
(Area under the curve of the Receiver operating characteristic) Performance measurement for classification problems that measures how well a model separates categories. An AUC score close to 1 means that the model is good at separating categories; it classifies 0s as 0s and 1s as 1s. An AUC score close to 0 means that the model is separating into opposite categories; it classifies 0s as 1s and 1s as 0s. An AUC score close to 0.5 means that the model is no better at separating categories than random chance.
A table that is used to display the performance of a classification model on a test data set as compared to the ground truth. Each square has a unique row and column combination. The square in row 1 and column 2 says that the ground truth classified those test data points as category 1 but the classification model classified those same test data points as category 2. All of the squares along the diagonal of the table display the number of test cases when the ground truth and classification model are in agreement. All squares off the diagonal represent any disagreement between the ground truth and classification model; the further away from the diagonal, the larger the disagreement. If there is complete agreement between the ground truth and classification model, then all of the off diagonal squares will be zero. If there is complete disagreement between the ground truth and classification model, then all of the diagonal squares will be zero.
If you experience any issues during the challenge, please let us know by emailing ailabsupport@acr.org.
Welcome to the ACR AI-LAB™ BREAST DENSITY CROWDSOURCING CONTEST. Annotate as many breast density scans as you can over the course of the ACR's annual meeting.
Select the use case for which you would like to prepare the data for.
Select the data set that contains the images that you would like to prepate.
If no data sets are listed, you will need to create them first in Data Manager.
Select to convert the Dicom images on the dataset into png images for annotation purposes.
Select to convert the Dicom images in the dataset into Numpy objects for model creation.
Welcome to the ACR AI-LAB™ BREAST DENSITY CROWDSOURCING CONTEST. Annotate as many breast density scans as you can over the course of the ACR's annual meeting.
Welcome to the ACR AI-LAB™ Residency Mammography Challenge.
