On Training Model Bias of Deep Learning based Super-resolution Frameworks for Magnetic Resonance Imaging

Mamata Shrestha; Ukash Nakarmi; Nian Wang

Abstract

Super-resolution is an important technique in various fields, particularly in medical imaging, where it plays a crucial role in understanding and analysis of complex, qualitative, and quantitative characteristics of tissues at high resolutions. However, obtaining high-resolution images often faces practical limitations pertaining to acquisition device limitations, patient motion, or longer acquisition times. The remarkable success of deep learning methods has recently opened doors to their application in image super-resolution tasks as well. These deep learning-based methods heavily rely on a substantial amount of data, which is often unavailable, especially in the case of Magnetic Resonance Imaging (MRI) scans. Particularly in magnetic Resonance super-resolution, it is often impossible to have low-resolution and high-resolution training image pairs. To address this, deep learning approaches simulate low-resolution images using many image degradation methods mimicking low-resolution images to create training image pairs from the available few high-resolution images. However, models trained on specific degradation simulations exhibit bias, leading to poor performance in real-world scenarios. In this paper, we hypothesize that such deep learning models trained on specific training image pairs with a specific degradation model are biased, we systematically study such biases with different types of degradation, different deep learning frameworks, and training losses. Finally, we advocate ensuring the diversity of degradation models to generate training image pairs controls such biases resulting in a more robust learning framework for MR image super-resolution.

Keywords: nan

Links
[Full text PDF][Bibtex]