Skip to content
Snippets Groups Projects
Commit d37fa77b authored by Tu, Ethan's avatar Tu, Ethan
Browse files

Upload New Machine Learning Report 3/13

parent cfe77815
No related branches found
No related tags found
No related merge requests found
%% Cell type:markdown id: tags:
# <center>Using Machine Learning (ML) in Perfusion Kinetics Modeling</center>
<center>by Ethan Tu</center>
%% Cell type:markdown id: tags:
Dynamic Contrast-Enhanced (DCE) imaging techniques aim to assess and characterize vasculature of the body by relating tracer-perfusion kinetic models to measured concentrations by MRI or CT. Machine learning is a great tool that is already being used in this field. For example, in a model of a tracer perfusing into surrounding tissue, we have four major parameters that need to be optimized; flow, surface area capillary constant, volume of the interstitial fluid, and baseline concentration in the tissue. Machine learning has already been extensively used in optimization problems and this area is no different. However, since I have already discussed this problem in the Optimizations report, I will instead discuss machine learning in relation to perfusion kinetics through estimation of parameters through supervised learning of DCE images.
Deep learning (DL) in DCE images have typically been used as an image classifier or for segmentation. Given a MRI or CT, does this patient have the disease or not? Given an image, can DL distinguish one area of the body from another? These have been the most important questions being solved by DL. Combining the two, we can ask ourselves: Given an MRI or CT scan, can DL distinguish blood plasma from surrounding tissue and can it estimate the amount of tracer in both areas? Using convolutional neural networks and supervised learning, we can answer such a question.
There are many aspects to solving this issue. For example, Mckinely et al., studied model performance of CNNs with respect to increasing sample size and voxel patch sizes to estimate arterial input functions.$^1$ Optimal sample size to train the network was estimated to be around 5000 images with a voxel patch size of 5x5. A whole new field of radiomics has been developed to turn medical images into mineable high-dimensional data by extracting a high number of handcrafted quantitative imaging features based on a wide range of mathematical and statistical methods.$^2$ DL has also reduced reconstruction time of MRI and CTs, which have the potential to speed up the model development process.$^3$ Most importantly, PK parameters can be estimated using DL techniques. Ulas et al. trained a CNN model to yield PK parameters which can better discriminate different brain tissues, including stroke regions, given DCE-MRI images.$^4$ Normally, DCE-MRI suffers from poor spatial resolution, noise, and insufficient volume coverage. Therefore, to train their CNN, they used a custom loss function that incorporated a feed-forward physical model that relates the PK parameters to corrupted image-time series obtained due to subsampling in k-space.$^5$ In laymans' terms, they trained their CNN taking into account for a bunch of different types of artifacts and poor image quality. Because of this, their model generalizes well to new cases even if a subject specific arterial input function (AIF) is not available for the new data.
Combining all the research together has yet to be done. If we can improve the estimation of the arterial input function, we can improve the overall estimation of PK parameters. If we can improve feature extraction or image data, we can improve estimation of PK parameters. If we come up with a better neural network with optimal weights, we can improve estimation of PK parameters. There is a bunch of different methods of solving this problem, and hopefully they can all be combined in the future.
%% Cell type:markdown id: tags:
---
# References
[1] A. S. Lundervold and A. Lundervold, “An overview of deep learning in medical imaging focusing on MRI,” Zeitschrift für Medizinische Physik, vol. 29, no. 2, pp. 102–127, May 2019, doi: 10.1016/j.zemedi.2018.11.002.
[2] R. McKinley, F. Hung, R. Wiest, D. S. Liebeskind, and F. Scalzo, “A Machine Learning Approach to Perfusion Imaging With Dynamic Susceptibility Contrast MR,” Front Neurol, vol. 9, Sep. 2018, doi: 10.3389/fneur.2018.00717.
[3] T. Leiner et al., “Machine learning in cardiovascular magnetic resonance: basic concepts and applications,” Journal of Cardiovascular Magnetic Resonance, vol. 21, no. 1, p. 61, Oct. 2019, doi: 10.1186/s12968-019-0575-y.
[4] C. Ulas et al., “Convolutional Neural Networks for Direct Inference of Pharmacokinetic Parameters: Application to Stroke Dynamic Contrast-Enhanced MRI,” Front Neurol, vol. 9, Jan. 2019, doi: 10.3389/fneur.2018.01147.
[5] C. Ulas et al., “Direct Estimation of Pharmacokinetic Parameters from DCE-MRI using Deep CNN with Forward Physical Model Loss,” arXiv:1804.02745 [cs], Jun. 2018.
0% Loading or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment