Preliminary Results: Comparison of Convolutional Neural Network Architectures as an Auxiliary Clinical Tool Applied to Screening Mammography in Mexican Women

Purpose

Mammography is the modality of choice for the early detection of breast cancer. Deep learning, using convolutional neural networks (CNNs) specifically, have achieved extraordinary results in the classification of diseases, including breast cancer, on imaging. The images used to train a CNN varies based on several factors, such as imaging technique, imaging equipment, and study population; these factors significantly affect the accuracy of the CNN models. The aim of this study was to develop a novel CNN for the classification of mammograms as benign or malignant and to compare its utility to that of popular pre-trained CNNs in the literature using transfer learning. All CNNs were trained to detect breast cancer on mammograms using mammograms from a created database of Mexican women (MAMMOMX-PABIOM) and from a public database of UK women (MIAS).

Methods

A database (MAMMOMX-PABIOM) was built comprising 1,070 mammography images of 235 Mexican patients from 4 hospitals in Mexico. The study also used mammographic images from the Mammographic Image Analysis Society (MIAS) public database, which comprises mammography images from the UK National Breast Screening Programme. A novel CNN was developed and trained based on different configurations of training data; the accuracy of the models resulting from the novel CNN were compared with models resulting from more advanced pre-trained CNNs (DenseNet121, MobileNetV2, ResNet 50, VGG16) which were built using transfer learning.

Results

Of the models resulting from pre-trained CNNs using transfer learning, the model based on MobileNetV2 and training data from the MAMMOMX-PABIOM database achieved the highest validation accuracy of 70.10%. In comparison, the novel CNN, when trained with the data configuration A6, which comprises data from both the MAMMOMX-PABIOM database and the MIAS database, produced a much higher accuracy of 99.14%.

Conclusion

Although transfer learning is a widely used technique when training, data is scarce. The novel CNN produced much higher accuracy values across all configurations of training data compared to the accuracy values of pre-trained CNNs using transfer learning. In addition, this study addresses the gap in that neither a national database of mammograms of Mexican women exists, nor a deep learning tool for the classification of mammograms as benign or malignant that is focused on this population.