Visualisation and Analysis of Learned Features from Deep Convolutional Networks in Breast Cancer
Breast cancer is the most common invasive cancer in women. Worldwide, approx. 12 percent of women are a effected at least once in their lifetime. Mammograms are taken for regular screening and as one of the first diagnosis steps. After 10 annual screenings, around half of patients will receive at least one false positive test. The high incidence rate and low detection specificity eagerly demand for smarter computer-aided detection (CAD) systems that support radiologists to analyse mammograms and to audit their findings.
A method for tackling various computer vision problems like image classification and object localisation are convolutional neural networks (CNNs). A CNN learns task-relevant visual features itself given training data, a learning algorithm and a loss function instead of having a human handcrafting and applying them during preprocessing. They are strong predictors and hence appear to be a practical choice for CAD. One major issue of those systems is the lack of interpretability. Without extensive annotation efforts, they only output whether an image contains breast cancer or not. It is not comprehensible whether the model has focused on the correct criteria that experts have in mind when analysing images. If being unjustifiably trustful, false decisions can be made. On the patient side, it can i.a. lead to unnecessary anxiety and biopsies. It is hence desirable to make machine learning models explain themselves in terms that doctors understand and use at their work. In addition, new concepts might be explored through the data-driven approach. A promising way to address these problems is to visualise and analyse features within a CNN.