Breast Density in Mammography

Hi to all,

Does anyone knows of any app/code to classify mammograms in density terms? Or even an app that computes a density value.

I do not want to retrieve a segmentation, only a classification or a single final value.

Thank you!

Answer 1

[Aastha]

Hi @João Mendes,

I understand your objective is to regress a density value from a mammogram image. You may consider the following steps to accomplish this using a Convolutional Neural Network (CNN) in MATLAB:

1. Begin by preparing your dataset, where the input data consists of mammogram images and the corresponding target values are the associated density values. Store the image files in a folder and create a datastore or table that links each image to its label.

imageFolder = 'path_to_images'; 
imds = imageDatastore(imageFolder, 'FileExtensions', '.png', 'LabelSource', 'foldernames'); 

% Assuming labels are stored in a separate table or array 
load('densityLabels.mat'); % Contains variable `densityValues` 
tbl = table(imds.Files, densityValues, 'VariableNames', {'Image', 'Density'});

2. Next, define the CNN architecture using a layers array. The network should be structured for a regression task. You may use the following MATLAB code snippet for reference:

layers = [ 
imageInputLayer([M N 1]) 
convolution2dLayer(3,8,Padding="same") 
batchNormalizationLayer 
reluLayer 
fullyConnectedLayer(numResponses) 
]; 

3. Configure the training process by specifying training options with the "trainingOptions" function. Then, train the CNN model using the "trainnet" function. Choose an appropriate loss function for your regression task as per your requirements.

For more information on the functions mentioned above, you may refer to the following documentation pages:

• imageDatastore: https://www.mathworks.com/help/matlab/ref/imagedatastore.html
• List of Supported Layers: https://www.mathworks.com/help/deeplearning/ug/list-of-deep-learning-layers.html
• trainnet: https://www.mathworks.com/help/deeplearning/ref/trainnet.html

I hope this helps!

[idris]

Hi Mendes!

If you're looking to classify mammograms by breast density (without segmentation), you can build a solution in MATLAB using pretrained CNNs like ResNet or DenseNet. Just extract features from the whole image and train a classifier (e.g., SVM or softmax layer) to predict BI-RADS density categories or a continuous score.

🛠️ Key steps:

• Preprocess mammograms (resize, normalize)
• Use activations() to extract deep features
• Train a classifier with fitcecoc() or trainNetwork()
• Output a single class label or density value

No segmentation needed: just whole-image classification. Let me know if you want sample code or dataset suggestions!