CLSSS: Deep Learning for Image Reconstruction in Medical Imaging

Deep learning has the promise to revolutionize the field of image reconstruction in medical imaging. For example, preliminary studies have shown that deep learning approaches could allow for a ten-fold increase in the speed of MRI acquisitions, or allow for x-ray CT imaging at half the conventional radiation dose without compromising image quality. However, the success of deep learning is predicated on access to a large, well-curated database of ground truth training images. In many MRI and CT imaging applications, ground truth training images are scarce or non-existent, which makes the extension of off-the-shelf deep learning solutions to these settings challenging. Furthermore, in clinical settings it is essential that learning-based reconstruction methods do not hallucinate or erase critical diagnostic features in the images (e.g., tumors). Finally, deep learning approaches must also be robust to confounding factors such as noise, poor calibration, or patient movement. This talk will discuss these and other limitations of current deep learning approaches for MRI and CT image reconstruction, and highlight some possible solutions.

Speaker: Greg Ongie is a postdoctoral researcher in the Department of Statistics.