The repeatability of IVIM with fuzzy clustering algorithm in liver imaging
Kaining Shi1, Ying Liu2, Yu Shi2, and Qiyong Guo2

1Imaging Systems Clinical Science, Philips Healthcare, Beijing, China, People's Republic of, 2Radiology department, Shengjing Hospital of China Medical University, Shenyang, China, People's Republic of

### Synopsis

Fussy clustering technique (FCM) has been combined with IVIM to increase the stability and reduce the post-processing time of the nonlinear curve fitting in IVIM. Another problem of the widely used bi-exponential IVIM model is its poor repeatability. This work is to assess the repeatability of IVIM with FCM between two scans in healthy liver by calculating the coefficient of variation and the 95% Bland–Altman limits of agreements. Results proved that FCM could improve the repeatability of IVIM, especially for the parameter D*, which was the most unstable among total 3 parameters.

### Introduction

Fussy clustering technique (FCM) can increase the stability and reduce the post-processing time of DCE-MRI,1,2 by sorting all voxels’ time-signal curves into several types. To address the same problems in the Intravoxel Incoherent Motion (IVIM) imaging, which is also sensitive to noise and time-consuming, the FCM has been combined with the nonlinear curve fitting of IVIM.3 However, another problem of the widely used bi-exponential IVIM model is its poor repeatability.4-5 The purpose of this work is to measure the repeatability of IVIM with FCM between two scans in healthy liver, compared with regular pixel by pixel post processing method.

### Method

3 female and 2 male young healthy volunteers (24-27 years old) were scanned by a 3T whole body scanner ( Ingenia, Philips Medical System, Best, The Netherlands). The single-shot spin echo DWI was scanned twice in one exam with following parameters: TE/TR 73/6000ms, FOV 360x360mm, acquisition matrix 128x189, 24 slices with the thickness of 7mm and 1mm gap, SENSE 3, NSA 2. 8 b values ( 0, 20, 50, 100, 200, 350, 500, 800) were used.

The post-process was performed on home-made software. The bi-exponential model6 was employed for the curve fitting: $$S(b)/S(0)=(1-f)\times\exp(-bD)+f\times\exp(-bD*)$$ Mask was drawn manually to restrict the curve fitting to be performed only in liver ( Fig.1a ). To investigate the effect of cluster numbers, FCM with 20 clusters and 40 clusters per slice were run respectively. 7 ROIs in total were drawn on the Couinaud Segments7 2, 3, 4, 5, 6, 7, 8 respectively for each liver (Fig.1). The repeatability of two scans was assessed by calculating the coefficient of variation (CV), and the 95% Bland–Altman limits of agreements (BA-LA) of IVIM derived parameters( D, D* and f) for each post-process method. The CV was also calculated for all segments.

### Result

Three methods’ results of one volunteer’s two scans are demonstrated in Fig.1. D and D* is more homogenous with FCM. The mean value, standard deviation and CV of two scans are listed in Table.1. Both two FCM methods have lower CV value for the D, compared to the normal method. D* has the highest CV value in all three methods, while FCM with 20 clusters has the lowest CV of D*. However, normal pixel-by- pixel method has the lowest CV of f. The result of BA-LA assessment is shown in Fig.2. The CV of each segment is listed in Table 2.

### Discussion

In this work, each ROI was saved and used in three methods of both two scans. Without obvious motion between two scans, this strategy could maintain the area and location of every ROI and got much better repeatability than previous work.4,5 D* had the worst repeatability, which is consistent with previous reports, while FCM with 20 clusters can improve it with the price of increased CV of f. However, since the CV of f is much smaller than that of D*and the increase of f’s CV is much smaller than the decrease of D*’ CV, especially for those regions with high CV of D*(such as segment 2 and 5), FCM with 20 clusters improved the total repeatability of IVIM.

### Acknowledgements

No acknowledgement found.

### References

1. J Li, et al. Med Phy 2009;36:3786-3794. 2. W. Chen, et al. Academic Radiology 2006;13:63-72. 3. K Shi, et al. ISMRM(2015) 2905. 4. A. Andreou. et al. Eur Radiol (2013) 23:428–434. 5. Suguru Kakite, et al. JMRI 41:149–156 (2015). 6. Yedaun Lee, et al. Radiology 2015;274(2):405-415; 7.Couinaud C. Paris: Pers Ed. 1989,26-39;

### Figures

Fig.1 IVIM parameters ( D, D* and f) of three methods (normal pixel based curve fitting, FCM with 20 clusters, FCM with 40 clusters) from one volunteer’s two scans( 1st on left, 2nd on right). Mask was shown with green lines on the DWI image with b=80. 7 ROIs for Couinaud Segments 2- 8 were demonstrated.

Fig.2 The 95% Bland–Altman limits of agreements (BA-LA) of IVIM derived parameters( D, D* and f) for each post process method.

Table 1: The mean value, standard deviation and CV for IVIM derived parameters (D, D* and f) of two scans.

Table 2: The CV of each segment for three methods (normal pixel based curve fitting, FCM with 20 clusters, FCM with 40 clusters).

Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)
3467