Early-stage breast cancers are very challenging for computer-aided detection (CAD) because they are small and often blend in with surrounding tissues. One reason for the current CAD limitations may be the lack of temporal analysis. A radiologist usually uses the current and prior mammograms side by side to evaluate changes over time. We propose a CAD method for breast cancer screening using a recurrent neural network (RNN), a convolutional neural network (CNN) with follow-up scans. First, mammographic images are examined by three cascading object detectors to detect suspicious cancerous regions. This is similar to generating a region proposal. Then all regional images (one channel) are scaled to 224×224×3 and fed to a pre-trained CNN (ResNet-50 model) to extract features. The image features are extracted from a registered prior scan, a current scan, and their difference image, each of which has a dimension of 2048 prior to the fully-connected layer. Finally the features from the three images are combined to train a RNN classifier. The RNN functions as a temporal analysis, which can factor in multiple follow-up scans. Our digital mammographic database includes 102 cancerous masses, architecture distortion, and 27 healthy subjects, each of which includes two scans: current (cancerous or healthy), and prior scan (healthy typically one year before). Our experimental results show that the performance of the proposed CAD method is very promising.
|