Keywords:
Breast, Artificial Intelligence, Computer applications, Mammography, Computer Applications-Detection, diagnosis, Diagnostic procedure, Decision analysis, Cancer, Image verification
Authors:
J. P. Pereira Fontes, M. A. Guevara Lopez; Guimarães/PT
DOI:
10.26044/ecr2019/C-2062
Results
Experimental Setup:
The models were trained using a stochastic gradient descent learner.
For each run,
the model was trained for a maximum number of 150 epochs,
using a strategy of early stopping when there are no improvements.
This experimental setup allowed us to verify the integrity of our system and find the best combination of classifiers giving us the best performance in classifying breast cancer tumours.
To verify the integrity of these models,
we carried out a 10-fold cross-validation method for each model.
Film Mammography Results:
The use of the default ImageNet weights (included with the model definition) did not give good results,
what can be explained by the number of classes that are mapped by these weights.
The diversity of classes in the ImageNet dataset did not help in this particular case,
which led to the results obtained.
Therefore,
the fine-tuning showed to be an important and necessary step for the success and the performance achieved by our new proposed method.
In table 1 we show the results obtained after training using the BCDR-F03 dataset [11].
Digital Mammography Results:
In the attempt to enhance the scientific knowledge,
the best model trained using film mammographies was also tested in a dataset of Full Field Digital Mammography (FFDM) images (BCDR-D01).
For this purpose,
we use the trained model based on the BCDR-F03 dataset (a lower resolution dataset of images) to classify the images of the BCDR-D01,
which as mentioned is a high-resolution FFDM dataset.
Results can be observed in table 2.