Diffusion MRI (dMRI) signal is sensitive to the tissue architecture at the microscopic scale. Modern machine learning and deep learning techniques can be used to learn the mapping between acquired dMRI signal and specific features of the tissue microstructure. However, experimental design and validation of training sets are essential for reliable supervised and semi-supervised learning and reproducibility and uncertainty of prediction are still open questions. This lecture provides the key concepts behind machine learning applications to dMRI signal analysis for tissue microstructure quantification and show the audience various techniques which have been recently used.
Researchers and clinicians who are interested in dMRI and use or plan to use machine learning for analysing dMRI data and obtain quantitative maps of features of tissue microstructure.
Following this lecture, the audience will understand:
This lecture will cover the following topics:
Diffusion MRI and tissue microstructure
In dMRI, the acquired signal depends on the displacement of the water molecules inside the tissue on the order of tens of milliseconds. The distance travelled during this time is on the scale of microns and is similar to cellular sizes in biological tissue. Therefore, the molecular displacement is affected by the presence of cellular membranes and their properties (e.g. size, shape, density, orientation, permeability, etc), which makes the acquired signal sensitive to the tissue microstructure. By analysing the measurements acquired with different experimental parameters, it is possible to infer some of these properties, with the ultimate goal of providing quantitative maps of tissue features and potentially define more specific biomarkers of the tissue state [1].
Over the last decades, a significant effort has been made to improve both the measurement protocols and data analysis tools for dMRI in order to have better sensitivity and specificity to different tissue characteristics, while keeping the acquisition time clinically feasible.
Current research on dMRI includes the development of methods for the entire imaging pipeline, from designing better acquisitions, image reconstruction, microstructure parameter mapping, super-resolution, tractography, parcellation; to the detection of lesions, monitoring treatment response, etc. [2].
While some of these topics have been treated in the previous lectures, this lecture focuses on microstructure parameters mapping. It provides a basic overview of popular dMRI applications for tissue microstructure quantification and discuss their limitations, with particular focus on the brain tissue.
Conventional methods for microstructure feature mapping are based on mathematical modelling. Usually, mathematical modelling exploits a set of assumptions or theoretical foundations to model some properties of the dMRI signal. In general, mathematical models are used when the process that needs to be described is partly or completely known. However, sometimes accurate mathematical models that express relationships between some tissue microstructural features and the dMRI signal are intractable [3]. In such situations, as well as when a more complex and largely unknown phenomenon is under examination, machine learning methods can be more adequate.
Mapping tissue microstructure features with machine learning: current trends
This core part of the lecture describes the current trends and main methods used in machine learning and deep learning regression for dMRI microstructure. The methods will be grouped according to the main methodological classes. Examples of experimental design and applications relevant to tissue microstructure characterisation using dMRI are provided.
Random Forest (RF): is an ensemble learning method and supervised learning algorithm which can be used for regression tasks. It works by constructing a number of decision trees (forest) during training, one to each bootstrap sample drawn from the full set of training data. Once trained, the model typically outputs the ‘mean’ prediction over the full set of trees in the forest. RF has been employed for instance to tackle the intractability for some of the mathematical models of microstructure dMRI [4,5], and for dMRI super-resolution [6].
Bayesian Models (BM): Bayesian modelling relies on Bayes' theorem for updating the probability of a hypothesis as more evidence is obtained. Bayesian modelling is especially useful when data is limited, allows the problem of overfitting to be avoided and is also useful as an approach for modelling uncertainty. For example, a supervised machine learning approach based on a BM estimator has been used to improve the estimate of microstructure features of brain tissue [7-9].
Artificial Neural Networks (ANN): an ANN is an ensemble of perceptrons, where a perceptron is defined by a non-linear transfer function f and by two set of parameters: W (the weights) and b (the bias). The output of each neuron is the linear combination of the input x with the W added to the bias b, followed by the application of the transfer function (i.e. sigmoid or hyperbolic tangent function). One of the most popular ANNs is the Multi-Layered Perceptrons (MLP) that organizes the neurons in many different layers. MLP has been successfully used in dMRI, for example, for voxel-wise microstructure parameter estimation using several biophysical models [10].
Deep Neural Networks (DNN): to extract more sophisticated features, many hidden layers can be added to an ANN, defining a deep architecture referred to as a DNN. DNN has been used in dMRI microstructure for signal augmentation in order to reduce acquisition time for advanced microstructure modelling [11,12], dMRI super-resolution and uncertainty [13].
Guidelines for machine learning design and training
Finally, we provide general guidelines for designing and training machine learning methods for regression. We focus on microstructure feature estimation from dMRI data and discuss the basics of how to:build the training and testing setstune the hyperparametersevaluate precision and accuracy of the predicted features
[1] Callaghan PT. Principles of Nuclear Magnetic Resonance. Oxford & New York: Oxford University Press, 1991
[2] Alexander, Daniel C., Tim B. Dyrby, Markus Nilsson, and Hui Zhang. "Imaging brain microstructure with diffusion MRI: practicality and applications." NMR in Biomedicine (2017): e3841.
[3] Novikov, Dmitry S., Valerij G. Kiselev, and Sune N. Jespersen. "On modeling." Magnetic resonance in medicine 79, no. 6 (2018): 3172-3193.
[4] Nedjati-Gilani, Gemma L., Torben Schneider, Matt G. Hall, Niamh Cawley, Ioana Hill, Olga Ciccarelli, Ivana Drobnjak, Claudia AM Gandini Wheeler-Kingshott, and Daniel C. Alexander. "Machine learning based compartment models with permeability for white matter microstructure imaging." NeuroImage 150 (2017): 119-135.
[5] Fick, Rutger HJ, Neda Sepasian, Marco Pizzolato, Andrada Ianus, and Rachid Deriche. "Assessing the feasibility of estimating axon diameter using diffusion models and machine learning." In 2017 IEEE 14th International Symposium on Biomedical Imaging (ISBI 2017), pp. 766-769. IEEE, 2017.
[6] Alexander, Daniel C., Darko Zikic, Aurobrata Ghosh, Ryutaro Tanno, Viktor Wottschel, Jiaying Zhang, Enrico Kaden et al. "Image quality transfer and applications in diffusion MRI." NeuroImage 152 (2017): 283-298.
[7] Reisert, Marco, Elias Kellner, Bibek Dhital, Juergen Hennig, and Valerij G. Kiselev. "Disentangling micro from mesostructure by diffusion MRI: a Bayesian approach." Neuroimage 147 (2017): 964-975.
[8] Collier, Quinten, Jelle Veraart, Ben Jeurissen, Floris Vanhevel, Pim Pullens, Paul M. Parizel, Arnold J. den Dekker, and Jan Sijbers. "Diffusion kurtosis imaging with free water elimination: A bayesian estimation approach." Magnetic resonance in medicine 80, no. 2 (2018): 802-813.
[9] Sjölund, Jens, Anders Eklund, Evren Özarslan, Magnus Herberthson, Maria Bånkestad, and Hans Knutsson. "Bayesian uncertainty quantification in linear models for diffusion MRI." NeuroImage 175 (2018): 272-285.
[10] Bertleff, Marco, Sebastian Domsch, Sebastian Weingärtner, Jascha Zapp, Kieran O'Brien, Markus Barth, and Lothar R. Schad. "Diffusion parameter mapping with the combined intravoxel incoherent motion and kurtosis model using artificial neural networks at 3 T." NMR in Biomedicine 30, no. 12 (2017): e3833
[11] Golkov, Vladimir, Alexey Dosovitskiy, Jonathan I. Sperl, Marion I. Menzel, Michael Czisch, Philipp Sämann, Thomas Brox, and Daniel Cremers. "Q-space deep learning: twelve-fold shorter and model-free diffusion MRI scans." IEEE transactions on medical imaging 35, no. 5 (2016): 1344-1351.
[12] Gibbons, Eric K., Kyler K. Hodgson, Akshay S. Chaudhari, Lorie G. Richards, Jennifer J. Majersik, Ganesh Adluru, and Edward VR DiBella. "Simultaneous NODDI and GFA parameter map generation from subsampled q‐space imaging using deep learning." Magnetic resonance in medicine (2018).
[13] Tanno, Ryutaro, Daniel E. Worrall, Aurobrata Ghosh, Enrico Kaden, Stamatios N. Sotiropoulos, Antonio Criminisi, and Daniel C. Alexander. "Bayesian image quality transfer with CNNs: exploring uncertainty in dMRI super-resolution." In International Conference on Medical Image Computing and Computer-Assisted Intervention, pp. 611-619. Springer, Cham, 2017.