4542
Reconstruction of Tailored Magnetic Resonance Fingerprinting Using Random Forest Approach1Dayananda Sagar Institutions, Bangalore, India, 2Memorial Sloan Kettering Cancer Center, New York, NY, United States, 3Institute of Surgical Technology and Biomechanics, University of Bern, Bern, Switzerland, 4Wipro-GE, bangalore, India, 5Magnetic Resonance Research Program, Columbia University, New York, NY, United States
Synopsis
Magnetic Resonance Fingerprinting is a new acquisition/reconstruction technique to obtain multi-parametric map. Tailored MRF has demonstrated the quantification of longer T2 components contrary to classical MRF. The supervised learning based approach model in the study does not require construction of the dictionary. Leave out one approach has been utilized as the approach for modeling the random forest approach. The dictionary approach is heavy on the computation that limits the MRF to get into the clinic.
Purpose
Magnetic Resonance Fingerprinting (MRF) is a framework developed to accelerate acquisition-reconstruction to obtain multi-parametric maps simultaneously1. MRF approach thus far has utilized dictionary-based approach to obtain the multi-parametric maps1. The primary goal of this study is to apply random forest approach to Tailored Magnetic Resonance Fingerprinting (TMRF)2. Contrary to previous approaches reconstruction does not need dictionary based tissue mapping and proposed method also reduces the computation timeMethods
Acquisition-Four human in-vivo healthy volunteers were scanned to obtain brain data on GE 1.5T signa scanner as part of institution approved study. The spiral read out time was 5ms with fixed echo time (TE) of 2.7ms. Spiral trajectory had 48 arms each arm constituted of 1280 points. In total acquisition of 49s, 720 images were acquired including Inversion time (TI) and delay between the blocks for magnetization recovery. Three blocks were utilized to optimize contrast for T1, PD and T2. Each block consists of 240 acquisitions and three such acquisitions were carried out (total of 720 acquisitions). Signal intensity of a gradient echo based sequence is more dependent on FA than TR. Thus required contrast was obtained by optimal choice of FA. TRs and FAs were independently designed and combined to form single sequence as depicted in figure 1. TRs were generated based on smoothly varying perlin noise and FAs were generated based on the equation (1) in ref2
Reconstruction- Initial estimates are got through sliding window algorithm3. Random forest reconstruction is cast as a multi- output extra- tree regression; it aims to establish a non-linear relationship between signal evolution and parametric maps4. The regression algorithm built an ensemble of 40 regression trees and slitting of the nodes was carried out based on variance reduction measure. Minimum sample size was used as a metric for splitting of nodes in the tree4. The leaf was assigned with a vector of intensity values once the tree had grown. The estimates of each tree were then aggregated by arithmetic average to yield a final prediction. A leave one out evaluation strategy was employed for numerical evaluation of the proposed TMRF-RF approach as summarized in figure 2. Median filter was utilized to smoothen the noise generated due to over fitting of the data. Normalized root mean squared error(NRMSE) was plotted to measure the error between the predicted maps and ground truth
Results:
Three different slice of the brain are shown in the figure 3. The second slice clearly shows the ventricles of the brain. The results of T1 map reveal that the contrast between the grey matter and white matter in the predicted values from random forest decreased, but the CSF is over-fitted. The T2 map shows that there is not much difference between the predicted values of the maps to that of the ground truth. The contrast between the grey matter and white matter in T2 map is preserved and most of the CSF is not over fitted. The edge information in the predicted T2 map has been lost due to the application of median filter. NRMSE between the predicted and ground truth is at acceptable range as shown in figure 4.Discussion and Conclusion:
The proposed approach has the ability to predict the values with lesser computation time and predict the values with no dictionary. Multi- output feature of the random forest approach employed can predict different values simultaneously. LOO approach used can be replaced with other approach for the study in case of more datasets are available for training. Larger the dataset base is more, the training also improves resulting in less error for prediction.Acknowledgements
1. This work was supported by Vision Group on Science and Technology (VGST), Govt. of Karnataka, Karnataka Fund for strengthening infrastructure(KFIST), GRD#333/2015
2. Department of Science and Technology (DST), Govt. of India under the program Technology Systems Development (TSD) for the project “Novel acquisition and reconstruction strategies to accelerate magnetic resonance imaging using compressed sensing”, No: DST/TSG/NTS/2013/100-G.
3. Department of Information Technology (DIT), Govt. of India for the project "Indigenous - Magnetic Resonance Imaging (I-MRI)- A national Mission"
References
1.Dan Ma et al. “Magnetic Resonance Fingerprinting” Nature 2013; 495 (7440): 187-92
2. Imam et. al. “Tailored Magnetic Resonance Fingerprinting: Optimizing acquisition schedule and intelligent reconstruction using a block approach” ISMRM workshop on MRF 2017
3. cao et. al ISMRM 2016 4. Geurts et al. “Extremely randomized trees”, Machine Learning, pp.3-42, vol 63, 2006.
4. Geurts et al. “Extremely randomized trees”, Machine Learning, pp.3-42, vol 63, 2006.
Figures
