MRF denoising with compressed sensing and adaptive filtering

Zhe Wang, Qinwei Zhang, Jing Yuan, Xiaogang Wang

Abstract

Magnetic Resonance Fingerprinting (MRF) is a revolutionary approach to quantitative MRI, yet it is hindered by undersampling artifacts and discrete dictionary resolution errors. This paper proposes a three-stage pipeline—CS-Tree8AF—that integrates Compressed Sensing (CS), Decision Tree-based error detection, and adaptive filtering. By identifying and correcting "mismatched" pixels caused by the resolution gap, the method achieves significant improvements in map quality, increasing the Peak Signal-to-Noise Ratio (PSNR) of $T_1$ maps by over 4 dB.

Introduction

Magnetic Resonance Fingerprinting (MRF) allows for the simultaneous estimation of multiple tissue properties in a single, rapid scan. However, two major hurdles limit its clinical utility: artifacts from k-space undersampling and "salt-and-pepper" noise arising from the "resolution gap" in the pre-calculated dictionary. When a patient's tissue properties fall between dictionary entries, the system often selects a distant, incorrect match, resulting in large, non-Gaussian errors that appear as speckles on the final maps.

Methods

The authors introduce a three-step denoising workflow to address these limitations:

  1. Compressed Sensing (CS) Reconstruction: CS is applied to raw k-space data to remove undersampling artifacts by leveraging the spatial sparsity of medical images before dictionary matching begins.
  2. Error Prediction via Decision Trees: This stage identifies pixels where the dictionary match is likely incorrect. The authors found that mismatched pixels exhibit distinct similarity profiles, often characterized by lower peak scores and multiple local maxima.
Similarity profiles across dictionary indices, showing the features used for error detection.

A decision tree is trained using an 8-dimensional feature vector derived from the top four similarity scores and their indices to classify pixels as "correct" or "error."

  1. Adaptive Filtering: Pixels flagged as errors are replaced using an adaptive filter that ignores the "bad" pixel's own value, preventing error propagation into surrounding tissue.

Results

The proposed CS-Tree8AF method was evaluated using simulated brain phantoms. The statistical difference between correct and incorrect pixel matches provides a robust basis for the decision tree classifier.

Visual error map (a) and histograms of similarity scores for correct (c) versus incorrect (d) pixels.

As shown in the histograms above, incorrect pixels (d) display a markedly different distribution of similarity scores compared to correct pixels (c). Quantitatively, the method improved the $T_1$ map PSNR from 24.3 dB to 28.4 dB and the SSIM from 0.731 to 0.881. Qualitative results confirm the removal of "speckle" noise while preserving sharp anatomical details that are often lost with standard Gaussian filtering.

Conclusion

The CS-Tree8AF pipeline effectively mitigates the primary sources of error in MRF. By treating similarity profiles as diagnostic signals, the method provides a targeted approach to denoising that outperforms standard filtering techniques. Future work will focus on clinical validation and the integration of more advanced patch-based filters.

Summary updated

BibTeX
@inproceedings{wang2014mrf,
  author = {Zhe Wang and Qinwei Zhang and Jing Yuan and Xiaogang Wang},
  title = {{MRF denoising with compressed sensing and adaptive filtering}},
  booktitle = {ISBI},
  year = {2014},
  doi = {10.1109/isbi.2014.6868009},
}
Figure