Introduction

Magnetic resonance fingerprinting (MRF) could be a key technology for reproducible radiomic analyses, as it generates highly reproducible quantitative maps.^1-3 Pattern-matching in MRF is a distinctive feature in terms of reproducibility, since the reconstructed maps are dependent on the dictionary design. The influence of dictionary design on radiomic features should be considered when using datasets containing maps generated from different dictionaries. However, to the best of our knowledge, this has not been investigated previously. In this study, we investigated the influence of the dictionary step size on MRF radiomic features and evaluated the stability of each feature with different dictionaries. In addition, we introduced an inverse quantization process that could mitigate the influence of dictionary differences on radiomic features.

Methods

MR fingerprinting acquisition and dictionary matching
An overview of the process is illustrated in Figure 1A. Ten volunteers were scanned with 3D MRF using a 3-T GE 750w scanner, and written consent was obtained prior to the scans. The 3D MRF sequence was based on steady-state free precession with spiral projection k-space trajectory.^4,5 The acquisition parameters of MRF were as follows: field-of-view, 200×200×200 mm; matrix size, 200×200×200; and acquisition time, 9 min 51 sec. The dictionaries were generated using the extended phase graph formalism, with T1 and T2 values ranging from 10 ms to 3000 ms and 20 ms to 1000 ms, respectively. The step size for each dictionary was defined as shown in Figure 2. Quantitative maps were obtained using a maximum inner product search,¹ and we considered only MRF T1 maps for simplicity.

Data postprocessing
A spherical volume of interest (VOI) of diameter 20 mm was placed randomly inside the skull (Figure 2), and 500 VOIs (10 subjects, 50 VOIs per subject) were used in the analysis. PyRadiomics (version 3.0)⁶ was used to extract first- and second-order features from the VOIs for each dataset. For simplicity, we limited our analysis of second-order features to those derived from gray level co-occurrence matrices (GLCM)⁷, the most commonly used second-order feature in related literature. We used a default bin width of 25, and no image resampling, image intensity normalization, or image filter was used.

Statistical analysis
The interclass correlation coefficient (ICC) was calculated for each feature as a measure of the agreement between dictionaries, to evaluate the stability of the features against differences in the dictionary step size. We used a two-way random-effects model of the ICC (unit: single, type: agreement). To evaluate the effects of different step sizes on the radiomic features, the percent relative change was calculated with respect to corresponding references. The radiomic features obtained from the densest dictionary were used as references, assuming that it contains the highest amount of information.

Adjusting quantitative values obtained with a sparse dictionary
Using a highly sparse dictionary results in a low quantitative resolution. For example, the sparsest dictionary in our study has only 30 steps for the T1 values in the range of 100 ms to 3000 ms. We hypothesized that increasing the quantization steps of the quantitative maps derived from a highly sparse dictionary would improve the agreement between the features obtained with a highly sparse and dense dictionary. Patel et al. reported that a simple 3×3 spatial averaging filter can increase the dynamic range of images.⁸ However, this approach is not optimal, as it treats the center and neighboring voxels equally. Hence, we applied an inverse quantization process^8,9 using a Gaussian function,¹⁰ assuming that the voxel initially at the center has the most influence on the voxel value (overview shown in Figure 1B). A spatial deviation of 0.4 mm was chosen empirically.

Results

Effect of dictionary step size on texture features
Among the first-order features, uniformity and entropy were vulnerable to the dictionary step size (defined as ICC<0.80) (Figure 3). Other first-order features exhibited reproducible results for different dictionaries. In general, first-order features were more reproducible than GLCM features (Figure 4).

Adjusting quantitative values obtained with a sparse dictionary
The features calculated after the inverse quantization process exhibited a lower percent relative difference than those calculated before (Figure 5). Although some ICCs decreased after inverse quantization, the mean value improved from 0.55 ± 0.61 to 0.71 ± 0.23 after inverse quantization.

Discussion

Our results show that even when the same source data are used, different MRF dictionaries can yield different radiomic features. Therefore, images obtained with different dictionaries may affect radiomics analysis when used simultaneously. We introduced an inverse quantization process using a Gaussian function that mitigates the influence of using different dictionaries on the radiomic features. This indicates that by using an appropriate quantization process, images obtained with different dictionaries could possibly be utilized simultaneously in radiomic analysis.

Identifying conditions for which the differences between dictionaries are sufficiently smaller than clinically required precisions is a topic for future research. This may enable accurate radiomic analysis of MRF data from different dictionaries simultaneously.

Conclusion

Although visually indistinguishable, the radiomic features calculated from different maps may differ significantly. Therefore, special attention should be paid to maps obtained from different dictionaries during radiomic analyses.

Acknowledgements

No acknowledgement found.