Automatic mass detection in mammograms using deep convolutional neural networks.

Area of Science:

• Medical Imaging
• Deep Learning
• Artificial Intelligence in Healthcare

Background:

• Convolutional Neural Networks (CNNs) show promise in medical imaging analysis.
• Automated mass detection in mammograms is crucial for early breast cancer diagnosis.

Purpose of the Study:

• To propose and evaluate a patch-based CNN method for automated mass detection in full-field digital mammograms (FFDM).
• To investigate the effectiveness of transfer learning and domain adaptation for this task.

Main Methods:

• Evaluation of three CNN architectures (VGG16, ResNet50, InceptionV3) pre-trained on ImageNet.
• Training CNNs on the large CBIS-DDSM dataset and transferring the model to the smaller INbreast dataset for domain adaptation.
• Utilizing a fivefold cross-validation strategy and free-response operating characteristic (FROC) curves for performance evaluation.

Main Results:

• InceptionV3 demonstrated the best performance in classifying mass and non-mass regions on the CBIS-DDSM dataset.
• Transfer learning from CBIS-DDSM to INbreast significantly outperformed transfer learning from ImageNet, achieving a higher true positive rate (TPR) at fewer false positives per image (FPI).
• The proposed framework surpassed existing literature results on the INbreast database for mass detection.

Conclusions:

• Patch-based CNNs with domain adaptation offer a powerful approach for automated mass detection in mammography.
• Transfer learning from a large, digitized mammogram dataset (CBIS-DDSM) to a smaller, digital dataset (INbreast) is highly effective.
• The developed method shows significant potential for improving the accuracy and efficiency of breast cancer screening.