Healthcare & Bio AI
Deep-learning method classifies mammograms as benign or malignant breast cancer with accuracy above 96 percent, a 2025 study reports
In a 2025 study, researchers report that a deep-learning method classified mammograms as benign or malignant breast cancer with accuracy above 96 percent. The method processes digital mammograms, which are X-ray images of the breast. It compares several image-recognition model families called VGG and DenseNet and uses a SoftMax classifier, a method that turns model scores into class probabilities, to select the best setup. The researchers then reduced the learned features with a 50 percent dropout step and combined the remaining features before making a prediction. They checked performance with three-fold cross-validation, a way of rotating the data so each part is tested once.
How did the study test the method?
After collecting and resizing mammogram images, the researchers extracted features with the deep-learning models. They tested VGG and DenseNet variants, selected the best model using the SoftMax classifier, applied a 50 percent dropout step, and combined the features into one vector. They measured the final result with three-fold cross-validation.
What did the evaluation find?
The authors report accuracy above 96 percent for classifying mammograms as benign or malignant breast cancer. The abstract presents this as the result of the study's own evaluation and does not report a separate test on new patients or hospitals.
What questions come up?
What does accuracy above 96 percent mean here?
It means the method put more than 96 percent of the tested mammogram cases into the category the study treated as correct. The abstract does not say how many mammograms were tested or where the reference labels came from.
Was the method tested on new patients or hospitals?
The abstract does not report that. It reports three-fold cross-validation and does not report an independent test on a separate group of patients or hospitals.
Does this result mean the method is ready for hospital screening?
The abstract does not say that. It describes a computational evaluation, not a clinical deployment or a comparison with radiologists.
Limits and what's next
The abstract describes a computational evaluation. It does not name the dataset, the number of mammograms, or the setting where the images were collected. The reported accuracy is above 96 percent, and the abstract reports no independent test on a separate group of patients or hospitals. The abstract also does not report a comparison with radiologists or a test inside a real screening workflow. The abstract reports no independent replication, so the result is one study's reported performance.
Source: A. Santhi Mary Antony et al., “Breast Cancer Detection from Mammogram with Deep Features Fusion,” 2025.