Accelerating Water Removal in MRSI Using Domain Knowledge-Based Deep Learning

Magnetic resonance spectroscopic imaging (MRSI) is a non-invasive technique that provides spatially detailed information about metabolite levels. However, dominant water peaks can interfere with the detection of key metabolite signals. In this work, we propose an unsupervised deep learning model to enhance water peak removal in MRSI. Our approach incorporates domain knowledge into a deep autoencoder model to reconstruct and subtract water peaks, using both Gaussian and Lorentzian line shapes. Experimental results on an in vivo dataset demonstrate that our model performs comparable to the widely used HLSVD-PRO method, while significantly reducing computation time. A comparison of the mean squared errors within the water frequency range (4.0-6.0 ppm) shows a strong agreement between both methods in water removal, with an r2-score of 0.987. Furthermore, our model achieves a slightly higher water suppression ratio (27.35 vs. 20.19) and a lower metabolite reconstruction error (0.28 vs. 0.42). These findings suggest that the proposed model offers improved efficiency in water peak removal. The code for our model is available at https://github.com/Buchali/wr-mrsi.