Domain Adaptation of MRI Scanners

• Author / Creator

  Kushol, Rafsanjany

• Deep learning (DL) has become a leading subset of machine learning (ML) and has been successfully employed in diverse areas, ranging from natural language processing to medical image analysis. In medical imaging, researchers have progressively turned towards multi-center neuroimaging studies to address complex questions in neuroscience, leveraging larger sample sizes and aiming to enhance the accuracy of DL models. However, challenges arise due to variations in imaging characteristics across centers, often attributed to differences in MRI scanners. This phenomenon, known as domain shift, leads to inconsistent performance of DL models when applied to unknown test data. Domain adaptation (DA) methods aim to bridge this domain gap by aligning data across different domains. Unfortunately, the lack of suitable tools for domain shift analysis hinders the development and validation of DA techniques. Moreover, existing solutions often process entire datasets without accounting for source/target domain heterogeneity. Furthermore, the impact of various MRI scanners on different disease classification tasks remains largely unexplored.

  Motivated by the aforementioned challenges and limitations of existing literature, I first propose a novel framework called DSMRI (Domain Shift analyzer for MRI) to comprehensively assess the extent of domain shift within MRI datasets. This framework provides key insights into domain shift factors by integrating knowledge from diverse domains, including spatial, frequency, wavelet, and texture analysis. Secondly, I introduce another unsupervised framework called DeepDSMRI, which analyzes domain shift in MRI data using various deep models pre-trained on the ImageNet dataset. DeepDSMRI demonstrates its efficacy in determining domain shift not only in structural MRI (e.g., T1-weighted, T2-weighted, and FLAIR) but also in advanced MRI modalities such as diffusion-weighted imaging (DWI) and functional MRI (fMRI). To the best of my knowledge, this is the first work to analyze and quantify domain shift in multi-modal MRI using DL without requiring additional training on MRI data.

  Thirdly, I investigate the impact of scanner vendor variability on various disease classification tasks using multiple DL models. My analysis reveals a significant decline in classification accuracy when DL models are tested with data from different scanner manufacturers. To address the challenging task of amyotrophic lateral sclerosis (ALS) classification, where existing methods have not achieved satisfactory accuracy, I propose an effective and robust transformer-based framework called SF2Former. Leveraging the vision transformer (ViT) concept, SF2Former employs a novel linear fusion of spatial and frequency domain information to efficiently extract robust local and global discriminative features. This study pioneers in applying a transformer-based deep model for ALS classification, achieving state-of-the-art performance compared to existing popular ML methods.

  Finally, a new perspective in solving the domain shift issue for MRI data is designed by identifying and addressing the dominant factor causing heterogeneity within the dataset. An unsupervised DA method called DAMS (Domain Adaptation of MRI Scanners) is developed to align domain-invariant features between source and target domains by minimizing discrepancies in their feature maps. Instead of treating the entire dataset as a single source or target domain, the method processes data based on the primary factor driving variations. Furthermore, my research extends the concept of handling domain shift through black-box source-free domain adaptation (SFDA), which aggregates knowledge from multiple source domains and eliminates the need to access source data during target domain adaptation. This thesis offers innovative solutions to domain shift challenges in MRI data analysis, benefiting researchers not only in medical imaging fields but also in computer vision.

• Subjects / Keywords

  • Domain adaptation
  • Deep learning
  • MRI
  • Domain shift

• Graduation date

  Fall 2024

• Type of Item

  Thesis

• Degree

  Doctor of Philosophy

• DOI

  https://doi.org/10.7939/r3-aq6e-8y46

• License

  This thesis is made available by the University of Alberta Library with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.