Synopsis

Keywords: Image acquisition: Machine learning

The collaboration between academia and industry is vital for advancing the research, translation, and practical implementation of new deep learning techniques for MR reconstruction. There are multiple ways in which academia can contribute, which will be discussed in this lecture.

In the context of MRI, image reconstruction plays a crucial role in extracting meaningful information from the acquired data. Advanced reconstruction methods providing high quality images from undersampled data have been employed for years and are changing the way MRI is used in the clinic. Deep learning-based reconstructions have given another push to fast imaging and in improving image quality, and such methods did very quickly find their way in commercial products. However, there are still many opportunities for deep learning to further improve the image quality and robustness as part of image reconstruction.
The collaboration between academia and industry is vital for advancing the research, translation, and practical implementation of new deep learning techniques for MR reconstruction. There are multiple ways in which academia can contribute, which will be discussed in this lecture.


1. Robustness and generalization
Deep learning models for MRI reconstruction should be able to handle diverse imaging scenarios such as different organs, acquisition protocols, variability in patient size, heart rate, or presence of pathology. Deep learning reconstruction models developed for a very narrow application may show improved performance, but have limited practical applicability.
Academia can contribute by developing novel network architectures and training strategies that enhance robustness and generalization.
2. Computational efficiency
The deployment of deep learning models in a clinical setting requires fast image reconstruction that seamlessly integrates in the clinical workflow. In case of image guided therapy, real time or near real time reconstructions are needed. This typically involves efficient software implementation and computational hardware. Academia can contribute by developing fast
deep learning reconstruction techniques or data compression techniques enabling fast data transfer and short reconstruction latency in case of remote reconstruction.
3. Interpretable and explainable models
Deep learning methods are often criticized for their lack of transparency and interpretability. It is crucial to develop techniques that can provide insights into the decision making process of the models and generate confidence measures. Academia can contribute by designing explainable AI methods specifically tailored to MRI reconstruction tasks.
4. Training without available ground truth
Acquiring clinical data for training deep learning based MR reconstruction is a slow and expensive process. Acquiring fully sampled ground truth data may not be just inconvenient but in some cases simply impossible due to factors such as physiological motion. Academia can contribute by developing methods that don’t require a large clinical datasets as well as developing methods for realistic data simulations that can generate training data that is difficult or impossible to acquire.
5. Automatic image quality assessment
Developing deep learning MR reconstruction approaches requires very rigorous clinical validation. The ability of AI models to generate photorealistic images raises the justified concern that AI-based image reconstruction may cause artifacts that are not easily detected potentially causing misdiagnosis. Current image quality assessment metrics largely fail to capture subtle or local image degradation that can be detected by a trained observer and sometimes only a comparison to a high quality reference image can reveal small structural changes. Academia can contribute in developing methods for automatic image quality assessment that better resemble the assessment of a trained observer.
6. Clinical translation
Many tools reach the clinics but are not utilized for various reasons. In case of deep learning reconstruction this may be due to concerns about image quality or ease of use, but sometimes also the added value may also not be clear. Does the improved image quality lead to a measurable result like shorter reading time or more accurate diagnosis? What is the potential impact of the reduced scan time - is it reduced waiting times, larger number of patients served, or does this maybe also lead to faster time to therapy? Clinical and economic impact assessment of the new deep learning techniques are going to be crucial for their adoption and academia can contribute by conducting such investigations.
7. Addressing unmet image quality needs
Reconstruction from undersampled data is just one way deep learning can be used for MRI reconstruction. Image enhancement is another aspect that has received a lot of attention. However, reducing or removing various image artifacts may be the next step in deep learning reconstruction. Combined with MR physics models one could achieve more robust and reproducible image quality. Motion remains one of the most important issues even with the drastically reduced scan times. Image translation or synthesis can help to reduce image artifacts or replace a scan if it was too much corrupted. Quantitative imaging can be made more reproducible by using more comprehensive signal models and deep learning can help robustly estimating the parameter maps using these models. Academia can contribute in developing such models further pushing the limit of image quality and reproducibility.

Acknowledgements

No acknowledgement found.

References

No reference found.