The post processing - by that we mean the image processing of a medical ultrasound image - is of utmost importance for the final visual image presented to the clinicians using an ultrasound scanner. Years of experience and engineering goes into fine tuning the post processing to achieve the best possible visual look of the images. Recently, is has been shown that the entire image post-processing step can be learned by a deep neural network. More specifically using generative adversial networks. This has been demonstrated by Dietrichson, Smistad et. al. at NTNU, and recently the actual proprietary post-processing of a Simens Ultrasound Scanner was learned by a MimicNet and presented in a paper by Huang et. al from Duke University. The trained MimicNet and example data is available at https://github.com/Ouwen/MimickNet.
The specific task of this master project is to modify the code to be used on ultrasound images processed with the UltraSound ToolBox (http://www.USTB.no). Other post processing algorithms can be compared with the deep learning post processing. It is also interesting to evaluate how different beamforming algorithms prior to the post processing affects the final image.
Qualifications:
- Excellent MATLAB programming skills (Object oriented MATLAB)
- Signal processing (IN3190/4190, INF4480, IN5450, and preferably IN3015/4015)
- Image processing/analysis (One or both of: IN2310, IN5520)
- Git Version Control
Litterature:
- Dietrichson, F., Smistad, E., Østvik, A., & Lovstakken, L. (2018). Ultrasound Speckle Reduction Using Generative Adversial Networks. IEEE International Ultrasonics Symposium, IUS, 2018-October, 22–25. https://doi.org/10.1109/ULTSYM.2018.8579764
- Huang, O., Long, W., Bottenus, N., Lerendegui, M., Trahey, G. E., Farsiu, S., & Palmeri, M. L. (2020). MimickNet, Mimicking Clinical Image Post- Processing under Black-Box Constraints. IEEE Transactions on Medical Imaging, 39(6), 2277–2286. https://doi.org/10.1109/TMI.2020.2970867
