Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Learning-based Active Noise Cancellation for MRI
The aim of this project is to investigate the feasibility of machine learning based active noise cancellation to improve patient comfort during MRI scans. Eventually, we will build a MRI compatible prototype of noise cancelling headphones.
Keywords: MRI, Magnetic Resonance Imaging, MATLAB, Machine Learning, Open Source, Python, Control Engineering, System Identification.
During Magnetic Resonance Imaging (MRI), we use sequences of rapidly switching magnetic field gradients in order to manipulate the spins in the subject. Due to the Lorentz force, the gradient coils exert vibrations on their housing. This vibration results in acoustic noise which can be the main source of patient discomfort. Patient comfort can be one of the limiting factors for maximum duration of the MRI exam and the quality of the resulting images. An overview of noise in MRI is given in [1]. fMRI sequences, for example, are used clinically on a regular basis, however, they can produce Sound Pressure Levels (SPL) up to 133dB at 3T[2]. Due to the high SPL inside the bore, patients are advised to wear double ear protection (ear plugs + isolating passive headphones).
The prediction of SPL using MRI sequence parameters (Audio Transfer Function, ARF) would enable us to predict the noise level during an examination and even adjust the MRI sequence accordingly to reduce noise in the future. It has been shown that the ARF complies to linear system theory[2] and prediction is possible up to limited accuracy.
Nowadays, Active Noise Cancellation (ANC) has become a standard in consumer headphones. ANC is typically based on a simple least mean squares fit (LMS)[3], where the algorithm needs a certain amount of samples to adjust to the noise source and filter effectively. If the noise characteristics change, the algorithm adjusts in a delayed fashion. In addition to this disadvantage, these products are not MR compatible, even at low field strengths below 1.5T. Employing a machine learning network and non-magnetic components, we expect that it will be possible to overcome the limitations of the linear LMS fit and reduce delay time in noise cancellation.
**References**
[1] McJury M, Shellock FG. Auditory noise associated with MR procedures: A review. J. Magn. Reson. Imaging 2000;12:37–45 doi: 10.1002/1522-2586(200007)12:1<37::AID-JMRI5>3.0.CO;2-I.
[2] Moelker A, Pattynama PMT. Acoustic Noise Concerns in Functional Magnetic Resonance Imaging. Hum. Brain Mapp. 2003;20:123–141 doi: 10.1002/hbm.10134.
[3] Ardekani IT, Abdulla WH. FxLMS-based active noise control: A quick review. APSIPA ASC 2011 - Asia-Pacific Signal Inf. Process. Assoc. Annu. Summit Conf. 2011 2011;1:839–848.
During Magnetic Resonance Imaging (MRI), we use sequences of rapidly switching magnetic field gradients in order to manipulate the spins in the subject. Due to the Lorentz force, the gradient coils exert vibrations on their housing. This vibration results in acoustic noise which can be the main source of patient discomfort. Patient comfort can be one of the limiting factors for maximum duration of the MRI exam and the quality of the resulting images. An overview of noise in MRI is given in [1]. fMRI sequences, for example, are used clinically on a regular basis, however, they can produce Sound Pressure Levels (SPL) up to 133dB at 3T[2]. Due to the high SPL inside the bore, patients are advised to wear double ear protection (ear plugs + isolating passive headphones).
The prediction of SPL using MRI sequence parameters (Audio Transfer Function, ARF) would enable us to predict the noise level during an examination and even adjust the MRI sequence accordingly to reduce noise in the future. It has been shown that the ARF complies to linear system theory[2] and prediction is possible up to limited accuracy.
Nowadays, Active Noise Cancellation (ANC) has become a standard in consumer headphones. ANC is typically based on a simple least mean squares fit (LMS)[3], where the algorithm needs a certain amount of samples to adjust to the noise source and filter effectively. If the noise characteristics change, the algorithm adjusts in a delayed fashion. In addition to this disadvantage, these products are not MR compatible, even at low field strengths below 1.5T. Employing a machine learning network and non-magnetic components, we expect that it will be possible to overcome the limitations of the linear LMS fit and reduce delay time in noise cancellation.
**References** [1] McJury M, Shellock FG. Auditory noise associated with MR procedures: A review. J. Magn. Reson. Imaging 2000;12:37–45 doi: 10.1002/1522-2586(200007)12:1<37::AID-JMRI5>3.0.CO;2-I. [2] Moelker A, Pattynama PMT. Acoustic Noise Concerns in Functional Magnetic Resonance Imaging. Hum. Brain Mapp. 2003;20:123–141 doi: 10.1002/hbm.10134. [3] Ardekani IT, Abdulla WH. FxLMS-based active noise control: A quick review. APSIPA ASC 2011 - Asia-Pacific Signal Inf. Process. Assoc. Annu. Summit Conf. 2011 2011;1:839–848.
The aim of this project is threefold:
1. Investigate the feasibility and accuracy of SPL prediction based on sequence parameters only performing audio measurements at our experimental MRI scanner. The resulting ARF will be used to generate training data based on arbitrary MRI sequences.
2. Enhance existing ANC algorithms by replacing the LMS fitting procedure by a machine learning network. Evaluation of performance is done by using recorded noise from a MRI measurement and comparing to existing ANC algorithms.
3. Build a prototype ANC headphone using MRI compatible components (microphone, pneumatic speaker), deploy the ML network and review its performance during an arbitrary imaging sequence.
Depending on the corresponding thesis (bachelor, master, semester) and the student’s interests, we can focus on several bullet points. Students with interest in control engineering and machine learning are well suited for this work.
The aim of this project is threefold:
1. Investigate the feasibility and accuracy of SPL prediction based on sequence parameters only performing audio measurements at our experimental MRI scanner. The resulting ARF will be used to generate training data based on arbitrary MRI sequences.
2. Enhance existing ANC algorithms by replacing the LMS fitting procedure by a machine learning network. Evaluation of performance is done by using recorded noise from a MRI measurement and comparing to existing ANC algorithms.
3. Build a prototype ANC headphone using MRI compatible components (microphone, pneumatic speaker), deploy the ML network and review its performance during an arbitrary imaging sequence.
Depending on the corresponding thesis (bachelor, master, semester) and the student’s interests, we can focus on several bullet points. Students with interest in control engineering and machine learning are well suited for this work.
Please contact dillinger@biomed.ee.ethz.ch if you are interested in the topic. Interested students are asked to send a CV and a transcript of records. For students from abroad, we also need to know how they plan to organize funding for their stay.
Please contact dillinger@biomed.ee.ethz.ch if you are interested in the topic. Interested students are asked to send a CV and a transcript of records. For students from abroad, we also need to know how they plan to organize funding for their stay.