Deep Siamese Network for Metric Learning on Chest X-Ray Data
Deep networks have demonstrated the potential for computer-aided diagnostics for decision support. They have been successful in imaging-based diagnostic tasks for a variety of diseases: neurological conditions such as Alzheimer’s disease, cancers, cardio-pulmonary diseases and many others. Inspired by these successes, this thesis focuses on the diagnosis of thoracic diseases from X-ray images using deep learning. Recently, deep convolutional neural networks have been successfully applied to X-ray data. However, this approach suffers from two drawbacks: (1) classification rates for certain diseases are insufficiently high enough for effective decision support, and (2) the disease features extracted are high-dimensional, limiting visualization of the disease space for knowledge discovery. To overcome these limitations, I explore deep metric learning to learn a distance function to measure the similarity between two X-rays. Specifically, I learn a Siamese network composed of two identical sub-networks that share the same architecture and weights. The Siamese network is trained using a contrastive loss function that aims to push dissimilar data points apart, while pulling similar data points close together. In addition to improving classification rates, this Siamese network also extracts lower-dimensional embeddings for visualization. The empirical evaluation demonstrates that Siamese networks outperform deep convolutional networks and are also able to extract visualizable embeddings.