| dc.description.abstract | Cancer is a prominent cause of premature mortality worldwide and often preventable through early detection. Magnetic resonance imaging (MRI) is a vital modality in the investigation of many types of cancers, due to its strong sensitivity to soft tissue contrast. Beyond this capability, MRI can provide quantitative or parametric information about biophysical tissue properties and microstructural processes, including molecular diffusion and tissue perfusion. Utilizing appropriate mathematical models for MRI data enables the extraction of these quantitative characteristics through a process known as model fitting.
Today, the conventional MRI approach for assessing tissue perfusion characteristics for cancer detection and treatment monitoring is dynamic contrast-enhanced MRI (DCE-MRI). However, the acquisition of DCE-MRI data necessitates the injection of a contrast agent, typically gadolinium. This approach may pose contraindications for individuals with renal insufficiency, as well as those who are pregnant or breastfeeding. Additionally, recent findings indicating gadolinium deposition in the brain raise concerns about the desirability of this approach.
As an adjunct to DCE-MRI, diffusion-weighted MRI (DWI) has been shown to provide increased specificity in cancer diagnosis. DWI utilizes the random motion of water molecules, also known as diffusion, to generate image contrast. By acquiring images at multiple diffusionweightings (b values), a diffusion model can be applied to the signal decay for each voxel. Fitting the classical mono-exponential model to the DWI data yields the apparent diffusion coefficient (ADC), a metric reflecting water diffusivity in the observed tissue. ADC has proven valuable in tumor detection, characterization, and assessing treatment response in oncological imaging.
In reality, the microcirculation of blood in the capillary network introduces a perfusion component alongside diffusion, contributing to signal decay in DWI, which has the potential to obviate the need for DCE-MRI. The biexponential intravoxel incoherent motion (IVIM) model takes this perfusion component into account. The IVIM model provides besides the diffusion coefficient also perfusion-related parameters such as the faster pseudo-diffusion coefficient D* (linked to capillary flow) and the perfusion fraction f. However, fitting the IVIM model to diffusion data is a challenging ill-posed inverse problem, primarily due to small perfusion fractions and the low signal-to-noise ratio (SNR) inherent to DWI. This leads to substantial inaccuracies in perfusion-related parameters when using conventional fitting approaches, impeding the clinical adoption of IVIM. Consequently, alternative fitting methods are necessary to facilitate the practical clinical implementation of IVIM.
Within this dynamic field of medical imaging, a transformative revolution has been sparked by the enhanced computing power experienced over the past decade, giving rise to modern methodologies and algorithms driven by artificial intelligence (AI) and deep learning. These cutting-edge techniques enable computers to discover complicated patterns in large data sets. In response to these advancements, this thesis aimed to investigate the application of deep learning for the generation of clear, detailed maps of perfusion-based biomarkers from DWI, with a primary emphasis on enhancing parameter estimation for the IVIM model within DWI. Within the context of deep learning parameter estimation, this thesis explored various methodologies, including diverse learning strategies, examination of various training and test data, exploration of network architectures, and optimization of hyperparameters.
This thesis is based on three papers. In Paper I, a prior unsupervised voxelwise deep learning IVIM fitting approach was refined by optimizing various network hyperparameters. The optimized approach successfully addressed unexpected correlations in the original suboptimal unsupervised approach. As a result, the optimized approach exhibited improved accuracy, independence, and consistency of both diffusion and perfusion-related parameters. Moreover, in simulations and in vivo data from pancreatic cancer patients, the optimized approach demonstrated superior performance compared to state-of-the-art IVIM fitting methods, where it showed the most detailed and significantly less noisy parameter maps. Notably, it excelled in detecting the most significant changes in IVIM parameters throughout chemoradiotherapy.
However, subsequent research towards Paper II revealed that the optimized unsupervised deep learning approach exhibited poor anatomy generalization when applied to the brain. Meanwhile, other research demonstrated that supervised deep learning approaches may exhibit training data bias, which also warranted investigation in the context of IVIM modeling. It is important to note that these supervised approaches are optimized utilizing parameter values as ground truth. However, ground truth parameters estimates are often lacking due to the ill-posed nature of the signal analysis problem in DWI. Consequently, these ground truth parameter estimates are synthetically generated by simulating parameter estimates according to a distribution, such as a uniform distribution, or alternatively, derived from conventional estimators applied to the original signal data.
Based on these findings, Paper II explored the impact of key training features, including the effect of training data and training length, on both unsupervised and supervised learning for voxelwise deep learning IVIM parameter estimation. The findings showed that extending training beyond early stopping could address parameter correlations and reduce errors, offering an alternative to the exhaustive hyperparameter optimization of Paper I. However, prolonged training resulted in increased sensitivity to noise for the unsupervised estimates, which resembled those obtained by least squares fitting. In contrast, supervised estimates displayed a higher precision, but also a notable bias towards the mean of the training distribution that could lead to potentially deceptive parameter maps. This led to the conclusion that while voxelwise deep-learning-based model fitting holds promise for IVIM parameter estimation, a careful evaluation of design choices and their impact on fitting performance and biases is essential.
While Paper I and Paper II extensively investigated voxelwise deep learning approaches that treat voxels as independent, it is essential to recognize that the tissue microenvironment is typically locally homogeneous. In these tissues, microenvironmental properties do not change randomly between adjacent voxels. Therefore, leveraging these potential correlations between relevant signals in neighboring voxels should enhance model-parameter fitting.
In Paper III, the investigation encompassed four sub-studies exploring means to incorporate such spatial information for deep learning parameter estimation in biophysical modeling, specifically applied to the IVIM model in DWI. This study revealed that training supervised on spatially-correlated synthetic data in patches effectively leverages spatial information with apparently reduced noise sensitivity, akin to signal averaging. This resulted in improved estimator accuracy and decreased inherent supervised bias. Notably, no noticeable improvements were observed for unsupervised learning. Moreover, attention models (transformers) outperformed convolution-based networks for this purpose. The recently proposed neighborhood-attention permitted training on larger receptive fields than conventional self-attention, leading to improved estimator performance. These improvements were quantitatively demonstrated in novel fractal-noise parameter maps that provided spatially correlated ground truth. Qualitative in vivo findings in brain DWI data were broadly comparable to the quantitative evaluations in simulations. Additionally, the study demonstrated further means to enhance the method by leveraging additional information from the test set, including underlying spatial variation and underlying parameter value distributions. These promising approaches have the potential to be extended to any biophysical model applied to signal data, extending beyond the scope of MRI.
In conclusion, this doctoral thesis has substantially advanced the field of deep learning diffusion parameter estimation. These contributions mark a substantial step towards the practical implementation of complex diffusion signal analysis problems, like IVIM, in clinical settings. | en_US |
