Convolutional Neural Networks for Breast Cancer Screening: Transfer Learning with Exponential Decay

In this paper, we propose a Computer Assisted Diagnosis (CAD) system based on a deep Convolutional Neural Network (CNN) model, to build an end-to-end learning process that classifies breast mass lesions. We investigate the impact that has transfer learning when large data is scarce, and explore the proper way to fine-tune the layers to learn features that are more specific to the new data. The proposed approach showed better performance compared to other proposals that classified the same dataset. 1 Background and objectives Breast cancer is the most common invasive disease among women [Siegel et al., 2014] Optimistically, an early diagnosis of the disease increases the chances of recovery dramatically and as such, makes the early detection crucial. Mammography is the recommended screening technique, but it is not enough, we also need the radiologist expertise to check the mammograms for lesions and give a diagnosis, which can be a very challenging task[Kerlikowske et al., 2000]. Radiologists often resort to biopsies and this ends up adding exorbitant expenses to an already burdened patient and health care system [Sickles, 1991]. We propose a Computer Assisted Diagnosis (CAD) system, based on a deep Convolutional Neural Network (CNN) model, designed to be used as a “second-opinion” to help the radiologist give more accurate diagnoses. Deep Learning requires large datasets to train networks of a certain depth from scratch, which are lacking in the medical domain especially for breast cancer. Transfer learning proved to be efficient to deal with little data, even if the knowledge transfer is between two very different domains [Shin et al., 2016]. But still using the technique can be tricky, especially with medical datasets that tend to be unbalanced and limited. And when using the state-of-the art CNNs which are very deep, the models are highly inclined to suffer from overfitting even with the use of many tricks like data augmentation, regularization and dropout. The number of layers to fine-tune and the optimization strategy play a substantial role on the overall performance [Yosinski et al., 2014]. This raises few questions: • Is Transfer Learning really beneficial for this application? • How can we avoid overfitting with our small dataset ? • How much fine-tuning do we need? and what is the proper way to do it? 31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA. ar X iv :1 71 1. 10 75 2v 1 [ cs .C V ] 2 9 N ov 2 01 7 We investigate the proper way to perform transfer learning and fine-tuning, which will allow us to take advantage of the pre-trained weights and adapt them to our task of interest. We empirically analyze the impact of the fine-tuned fraction on the final results, and we propose to use an exponentially decaying learning rate to customize all the pre-trained weights from ImageNet[Deng et al., 2009] and make them more suited to our type of data. The best model can be used as a baseline to predict if a new “never-seen” breast mass lesion is benign or malignant.