Feature learning based on connectivity estimation for unbiased mammography mass classification

Breast cancer is the most commonly diagnosed female malignancy worldwide. Recent developments in deep convolutional neural networks have shown promising performance for breast cancer detection and classification. However, due to variations in appearance and small datasets, biased features can be learned by the networks in distinguishing malignant and benign instances. To investigate these aspects, we trained a densely connected convolutional network (DenseNet) to obtain representative features of breast tissue, selecting texture features representing different physical morphological representations as the network's inputs. Connectivity estimation, represented by a connection matrix, is proposed for feature learning. To make the network provide an unbiased prediction, we used k-nearest neighbors to find k training samples whose connection matrices are closest to the test case. When evaluated on OMI-DB we achieved improved diagnostic accuracy 73.89±2.89% compared with 71.35±2.66% for the initial CNN model, which showed a statistically significant difference (p=0.00036). The k training samples can provide visual explanations which are useful in understanding the model predictions and failures of the model.