Blumberg, Stefano B.; (2024) New Deep Learning Techniques for 3D MR Image Enhancement. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text Thesis-StefanoBlumberg-2024-07-22.pdf - Accepted Version Download (35MB) | Preview

Abstract

This thesis proposes novel deep learning techniques for image enhancement, for which the primary application is improving the quality of real-world human Magnetic Resonance Imaging (MRI) scans. Obtaining high-quality MR images at a reasonable cost and acquisition time will contribute to advancing medical diagnostics and patient care. Machine learning has proven successful in image enhancement and related tasks in computer vision. However, differences in data type and task specifics have hindered the translation of these techniques to the medical domain. This thesis, therefore, proposes novel deep learning techniques that address multi-channel, high-resolution 3D MR images and the specifics of the image enhancement task, which may consist of increasing image resolution, change of contrast, and increasing the number of image channels. The main methodological contributions are: (i) A technique to trade memory usage with computational time during deep learning training, thereby reducing the memory load of handling MR images. (ii) A strategy to combine images from multiple MRI scanning centers and protocols to improve image enhancement. (iii) A novel neural network layer for the upsampling operation used in the enhancement task. (iv) An approach to estimating what would be a far more costly, many-channeled from a relatively less expensive, few-channeled image.

Download activity - last month

Export as XMLJSONCSV

Download activity - last 12 months

Export as JSONXMLCSV

Downloads by country - last 12 months

Export as JSONXMLCSV

Archive Staff Only

View Item