Transfer Learning to Improve Breast Cancer Detection on Unannotated Screening Mammography

Breast cancer ranks in the top half of the division of cancers that prove to be fatal. The early detection of the disease is pivotal for the survival of patients. Recently, advances in deep learning have been used for the detection of breast cancer. These systems are often trained with sizeable medical imaging datasets of screening mammography prepared and annotated by expert radiologists. When developing such models, researchers face difficulty in obtaining the datasets required for training. The study presents a transfer-learning based approach that attempts to address the problems faced by existing breast cancer detection models trained with large annotated screening mammography datasets. Further, these models do not perform well on datasets without such annotations. The study attempts to solve these problems by using a multi-stage approach to train a breast cancer detection model. The initial stage involves training a Convolutional Neural Network (CNN) based model on a large publicly available dataset of annotated breast cancer mammography. The study then uses Transfer Learning to exploit the hierarchical learning properties of CNNs to fine-tune the model on a dataset of unannotated mammograms. An evaluation of the model at detecting breast cancer in unannotated screening mammography resulted in an accuracy of 75% and a reduced false-negative rate, marking an improvement over the benchmarked existing approach. The findings of the study show that transfer learning-based approaches improve the performance of an unannotated screening mammography breast cancer detector. The study also suggests the use of such systems as potential computer-aided diagnosis (CAD) tools.