Mono, bi- and tri-exponential diffusion MRI modelling for renal solid masses and comparison with histopathological findings

Purpose To compare diffusion tensor imaging (DTI), intravoxel incoherent motion (IVIM), and tri-exponential models of the diffusion magnetic resonance imaging (MRI) signal for the characterization of renal lesions in relationship to histopathological findings. Methods Sixteen patients planned to undergo nephrectomy for kidney tumour were scanned before surgery at 3 T magnetic resonance imaging (MRI), with T2-weighted imaging, DTI and diffusion weighted imaging (DWI) using ten b-values. DTI parameters (mean diffusivity [MD] and fractional anisotropy [FA]) were obtained by iterative weighted linear least squared fitting of the DTI data and bi-, and tri-exponential fit parameters (Dbi, fstar,and Dtri, ffast,finterm) using a nonlinear fit of the multiple b-value DWI data. Average parameters were calculated for regions of interest, selecting the lesions and healthy kidney tissue. Tumour type and specificities were determined after surgery by histological examination. Mean parameter values of healthy tissue and solid lesions were compared using a Wilcoxon-signed ranked test and MANOVA. Results Thirteen solid lesions (nine clear cell carcinomas, two papillary renal cell carcinoma, one haemangioma and one oncocytoma) and four cysts were included. The mean MD of solid lesions are significantly (p < 0.05) lower than healthy cortex and medulla, (1.94 ± 0.32*10− 3 mm2/s versus 2.16 ± 0.12*10− 3 mm2/s and 2.21 ± 0.14*10− 3 mm2/s, respectively) whereas ffast is significantly higher (7.30 ± 3.29% versus 4.14 ± 1.92% and 4.57 ± 1.74%) and finterm is significantly lower (18.7 ± 5.02% versus 28.8 ± 5.09% and 26.4 ± 6.65%). Diffusion coefficients were high (≥2.0*10− 3 mm2/s for MD, 1.90*10− 3 mm2/s for Dbi and 1.6*10− 3 mm2/s for Dtri) in cc-RCCs with cystic structures and/or haemorrhaging and low (≤1.80*10− 3 mm2/s for MD, 1.40*10− 3 mm2/s for Dbi and 1.05*10− 3 mm2/s for Dtri) in tumours with necrosis or sarcomatoid differentiation. Conclusion Parameters derived from a two- or three-component fit of the diffusion signal are sensitive to histopathological features of kidney lesions. Electronic supplementary material The online version of this article (10.1186/s40644-018-0178-0) contains supplementary material, which is available to authorized users.


Background
As a result of the increased use of abdominal imaging, more (asymptomatic) small (≤ 4 cm) renal masses are incidentally discovered. In a series of 173 patients only 58% of kidney tumours < 4 cm were malignant, whereas all kidney tumours > 7 cm were [1]. Hence, a substantial amount of incidentally discovered renal masses is not malignant [2][3][4]. The management of renal lesions includes radical or partial nephrectomy, minimal invasive ablative techniques or active surveillance. Because of concern for chronic kidney disease, nephron sparing surgery is preferred [5,6] but more importantly unnecessary surgery should be avoided. One way to realize this is by distinguishing between lesion types and reliably diagnosing benign tumour types, such as oncocytoma, prior to treatment [7]. However, with currently available clinical imaging modalities, benign renal masses are indistinguishable from malignant renal masses [4,8].
Many magnetic resonance imaging (MRI) techniques have been explored as methods to differentiate between benign and malignant renal lesions or between renal cell carcinoma (RCC) subtypes [4,[9][10][11]. One promising technique is diffusion-weighted imaging (DWI), which allows quantification of water motion in tissues without administration of exogenous contrast materials [12][13][14][15][16]. The apparent diffusion coefficient (ADC), derived from a mono-exponential model, is believed to reflect tissue cellularity as a higher tissue density will amount to more restricted diffusion, hence a lower diffusion value. However, ADC values for different subtypes may overlap, making determination of cut-off values to distinguish between benign and malignant solid renal masses problematic [17].
More complex models of diffusion, such as the diffusion tensor model (DTI) and the intravoxel incoherent motion model (IVIM), allow deriving additional information. DTI-derived parameters fractional anisotropy (FA) and mean diffusivity (MD) have been correlated with histological parameters such as cell density and nuclear grade [18]. The IVIM model is a bi-exponential model that includes molecular diffusion and microcirculation of blood in the capillary network ('pseudodiffusion') [19]. A combination of pseudodiffusion fraction f bi and the perfusion-free diffusion coefficient D bi from IVIM model is able to differentiate between renal tumour types [20,21]. Recently, the IVIM model was expanded to a threecomponent model by adding an additional component that accounts for intermediately fast water motion in the kidney [22,23]. The aim of this study is to compare parameters obtained from DTI, intravoxel incoherent motion (IVIM), and tri-exponential models of the diffusion signal of kidney lesions, for the characterization of renal lesions. Because tumours are usually not uniform and may consist of several areas with different structural patterns, we compare diffusion parameters with histopathological results.

Subjects
Approval of our institution's ethical committee was obtained for this prospective study and all subjects provided written informed consent. From March 2016 to May 2017, sixteen patients (11 male, age 65 (range 50-76) years old, 5 female, age 60 (range 48-72), total group: age 64 (range 48-76) years old) who had suspected kidney tumours and were planned to undergo radical or partial nephrectomy based on standard clinical diagnostic criteria were included. After including the first five consecutive patients, patients were also selected on tumour size (≤ 4 cm on radiologic examination) in order to increase chance of including benign solid lesions. After surgical resection of the tumour, kidney tumour type was determined according to the WHO classification of tumours of the urinary system [24] by histopathological examination of 2-μm-thick sections of formalin-fixed and paraffin-embedded tumour tissue blocks using haematoxylin-eosin (HE) staining.

Data processing
To enable accurate parameter fitting all scans were corrected for (breathing) motion before further processing. Due to differences in motion between the right and left kidneys, they were cropped and processed as separate data sets, as described previously [22]. All pre-processing was performed using diffusion imaging analysis package DTItools [github.com/mfroeling/DTITools] [25] and image registration toolbox Elastix [http://elastix.isi.uu.nl/] [26]. First, T 2 scans were processed to correct for slice by slice misalignment due to acquisition in multiple breath-holds using a rigid 2D registration algorithm after being resampled to 2 mm isotropic using a single interpolation method. Finally, all DWI data was corrected for breathing motion, by registering them to the unweighted volume using a rigid 2D b-spline registration algorithm after which the DWI data was registered to the reference T 2 scan using a 3D affine registration algorithm [22].

Parameter maps
From the DTI data the FA and MD were calculated using an iterative weighted linear least squares algorithm with outlier rejection using ExploreDTI [27]. From the IVIM data, bi-and tri-exponential diffusion decay parameters, i.e. the mean diffusion (D bi for bi-exponential and D tri for tri-exponential fitting), and the signal fraction attributed to pseudo-diffusion (f star for bi-exponential and f fast and f interm for tri-exponential fitting), were obtained by fitting a two and three-component model to the multiple b-value DWI data, as described previously [22,23]. To make a comparison between the DTI and IVIM data, the mean diffusion from a mono-exponential fit, D mono was also obtained.
Regions of interest (ROIs) to segment the tumour volumes were manually defined on the combined T 2 and DWI data by the principal researcher (S.v.B., 4 years of experience) in agreement with an experienced radiologists (C.K., 12 years of experience) using image segmentation toolbox ITK snap [28]. ROIs were placed inside the tumour, rather than following the contour, to limit the contribution of the signal from other tissue types due to partial volume effect or imprecise image registration. For comparison of tumour tissue with healthy kidney parenchyma, the cortex and medulla in the healthy contralateral kidneys were segmented using an automated algorithm as in [22]. The mean and standard deviation of the diffusion parameters were obtained for healthy cortex and medulla and lesion ROIs. Parameter maps MD, D bi , D tri , f star , f fast and f interm were obtained for visual comparison with T 2 and (if available) photographs of the gross appearance of the resected kidney tumours before histological examination.

Statistical analysis
All statistical tests were performed using SPSS (version 23.0. Armonk, NY: IBM Corp.). Healthy cortex and medulla were compared with all solid lesions using a Wilcoxon Signed Ranks test. The means of parameters MD, FA, D bi , D tri , f star , f fast and f interm in healthy cortex and medulla, different types of RCCs, cysts and benign solid lesions were compared using multivariate analysis of variance (MANOVA). Bonferroni correction was applied, and a p-value < 0.05 was considered significant for all statistical tests.

Subjects and scans
All patients were successfully scanned. One scan was removed before processing due to artefacts resulting from a lower-back implantation. After visual inspection following processing two other scans were removed, due to poor motion correction results and an error in the data. In the remaining thirteen scans, thirteen solid lesions (average size of maximum diameter, determined by histopathological  examination 3.85 cm, range 0.8-7.5 cm) and five fluid-filled cysts were found. Examples of the raw acquired data and the data after motion correction and image registration together with the ROI placement in one lesion are shown in Fig. 1.

Histological examination
Of the solid lesions, eleven were considered to be malignant (nine clear cell RCCs (cc-RCCs), two papillary cell RCCs (p-RCCs) and two were considered to be benign (one haemangioma of the kidney capsule and one oncocytoma). Of the nine cc-RCCs, only one had a homogeneous microstructural pattern consisting mainly of clear cells. Others had considerable amounts of necrosis, sarcomatoid differentiation, haemorrhaging or deviating growth patterns such as papillary, tubular or cystic growth. In Fig. 2, histopathological features of several kidney tumour tissues are displayed. Figure 3 shows MRI data (T 2 and DWI b = 0) parameters maps and (where available) photographs of the gross appearance of the resected kidney tumours for: a contralateral unaffected kidney (A-D), a case of cc-RCC (E-H) and the oncocytoma (I-L). In the Additional file 1: Figure S1 the gross appearance, MRI data and parameter maps of all tumour types are shown. The fractions of the diffusion components are shown as merged f-maps where f fast , f interm and f slow are colour coded red, blue and green, respectively. For the unaffected kidney, f fast (red) was high in those areas with a high blood flow (e.g. large blood vessels) whereas f interm (blue) was high in areas with free water (e.g. the pyelum). The diffusion coefficient (D tri ) was homogeneous throughout healthy kidney parenchyma. Upon visual examination, in the maps obtained in a cc-RCC, D tri was lower throughout the tumour, and f slow had a high contribution. For the oncocytoma, the lesion did not seem much different from normal kidney parenchyma, although f slow and D tri appeared higher.

Parameter maps
In the p-RCC, the merged f-maps showed a small contribution from f fast and a larger contribution from f slow . The D tri map indicated a low diffusion coefficient. In the cc-RCC with sarcomatoid differentiation and the cc-RCC with papillary growth the photographs and the diffusion parameter maps showed a more heterogeneous make-up, indicating a more complex tumour. In the cysts, f slow had a high contribution and D tri was high.
For the haemangioma, f slow and D tri appeared higher.

Parameter analysis
The mean and standard deviations of parameter values of grouped lesions and measurements of healthy cortex and medulla are given in Table 2. In Additional file 2: Figure S2 D mono is plotted against MD for each lesion. In Additional file 3: Table S1, the values for D mono are displayed together with the other diffusion coefficients MD, D bi , and D tri . MD and D mono have a similar order of magnitude for each (group of ) lesion, whereas D bi and D tri are structurally lower.
In Fig. 4, each lesion is represented by a dot in the scatter plot, plotting DTI parameters MD versus FA (Fig. 3a), IVIM parameters D bi versus f star (Fig. 3b), three-component parameters D tri versus f interm (Fig. 3c) and D tri versus f fast (Fig. 3d). Cysts were recognizable by a high MD, D bi and D tri, but they were more grouped in D tri versus f interm /f fast due to a consistently low f interm and f fast. In D bi versus f star and D tri versus f fast several general groupings were identifiable; the cysts are located in the lower right corner, the cc-RCCs in the middle and p-RCC (Fig. 2b for microscopic photograph) in the middle to the left. The oncocytoma is located above the cc-RCCs and the haemangioma between cc-RCCs and the cysts. However, one of the cc-RCCs is closely located to the oncocytoma due to a high value of f star and f fast . The p-RCCs and cc-RCC with sarcomatoid differentiation seemed to have a slightly lower D bi and D tri than most cc-RCCs, but one cc-RCC has a much lower diffusivity than all other lesions. Figures 5 and 6 show box plots for MD (Fig. 5a), D bi (Fig. 5b), D tri (Fig. 5c), FA (Fig. 6a), f star, (Fig. 6b), f fast (Fig. 6c) and f interm (Fig. 6d). The f interm and f fast show more pronounced differences and less overlap between individual lesions than f star. and compared to MD, both

Discussion
In this study, we have compared the DTI, IVIM and three-compartment models of the diffusion signal for the characterization of renal lesions. In the study population of sixteen patients who received radical or partial nephrectomy, two lesions were found benign upon histological examination. This indicates that resection was unnecessary and could have been prevented if these lesions were identified as non-malignant prior to surgery. Based on our results the haemangioma could potentially have been identified using diffusion-derived parameters, notably, a high D bi and D tri and a low f star and f fast . Previous studies applying DWI to evaluate kidney tumour type concluded that kidney lesions generally have a lower ADC than normal kidney tissue [13,17,29] and some studies also found a higher mean ADC value in benign lesions than in malignant lesions [15,30,31]. Studies typically find a higher ADC in oncocytomas than in RCCs [12-15, 17, 30-32] and therefore, ADC is likely to be a valuable parameter in evaluating tumour type. However, in itself it is not sufficient as a clinical index due to overlap in values between tumour tissue types [16,17,33].
In line with previous studies MD in this study was significantly lower in solid lesions than in healthy tissue, and RCCs had a lower MD than benign lesions and cysts. Differences in diffusivity between tissue types are usually attributed to differences in tissue cellularity, a higher cellularity will result in more restricted diffusion and therefore, more aggressive lesions are expected to present with a lower diffusivity [15,21]. Our study also showed similar results: cysts had the highest diffusion coefficients whereas lesions with a high degree of necrosis had low diffusivity, which we have assigned to an increase in with a diffusion-restricting elements, such as macromolecules, and disorganisation. In the two-component IVIM model, a fast moving component is separated from the diffusion signal, resulting in a lower diffusion coefficient D bi . The fraction f star of the diffusion signal that is attributed to fast moving water has been correlated to renal tumour vascularity [21]. As in this study, a lower MD and D bi for p-RCC than cc-RCC and similar values for f star in p-RCCs were found previously [21]. A tri-exponential fit of the diffusion signal in the kidney was previously shown to provide additional information on structures associated with pseudodiffusion by separating the fast from intermediate pseudodiffusion resulting in signal fractions f interm associated with a diffusion rate in the order of magnitude of free water and f fast associated with perfusion [22,23], and comparable values to this study for parameters derived from DTI, two-and three-compartment fitting for healthy cortex and medulla were reported [22]. MD was the only diffusion coefficient to be significantly different in healthy tissue from RCCs but within the tumour type groups the MD range was wider whereas the differences between tumour types were more pronounced in D bi and D tri, . Because diffusion coefficients D bi and D tri exclude the fraction of the diffusion signal that is attributed to fast water movement, they were lower but more precise than MD. Therefore, D bi and D tri better reflect tissue diffusion, making both these parameters more specific for tissue cellularity. However, none of the diffusion coefficients could be used to reliably distinguish between lesion types, as diffusion coefficients overlap. For a better contrast between lesion types, tissue cellularity (MD, D bi or D tri ) can be combined with a measure for vascularisation (f star or f fast ). For example, a combination of D bi and f star was previously shown to discriminate between renal tumour subtypes [21].
There was a large variation in diffusion parameter values between individual cc-RCC tumours due to tumour heterogeneity and differences in microstructural make-up. In addition, drawing statistical inferences from this study is limited due to the small number of cases and the sensitivity of the analysis to outliers. Therefore we have also analysed the results from individual lesions, associating histopathological characteristics with the interpretations of diffusion parameters outlined above. In this analysis, low diffusion coefficients were found in cc-RCCs with a high tissue density (due to extensive necrosis or sarcomatoid differentiation) whereas high diffusion coefficients were found in cc-RCCs with cystic structures. Additionally, tumours with a high perfusion rate are characterized by a high value of f star or f fast whereas tumours with a low perfusion rate, such as the haemangioma, cysts and the cc-RCCs with cystic structures have lower f star or f fast . Hence, these diffusion parameters seem to be indicative of histopathological features of kidney tumours. Although both f star and f fast seem to correlate to vascularisation, only the three-component parameters f fast and f interm show significant differences between different tissues whereas two-component parameter f star does not, showing that the tri-exponential model provides additional information over the bi-exponential model.
Because of the limited amount of cases it is impossible to formulate conclusions regarding the characterisation of kidney tumour type. Therefore, we have also analysed individual lesions relating our findings to histopathological details. The initial results from this analysis indicate that diffusion parameters are sensitive to histopathological features of kidney lesions, which is a first step towards non-invasive characterisation of these lesions prior to treatment. An improvement to this study would be to spatially correlate histopathology to parameter measurements and maps [34]. This would result in more specific validation of diffusion parameters, confirming the correlation between diffusion parameters to histopathological features of tissues, such as cellularity, perfusion and cystic structures. In addition, to establish what parameter values should be used to confidently distinguish between benign and malignant lesions and draw statistically significant conclusions, a larger study population should be included. However, since kidney tumour type is unknown before histopathological evaluation, researchers have no control over which tumour types are included. To increase the amount of benign kidney tumours, only small (≤ 4 cm) renal masses (about 40% benign [1]) can be included. Additionally, this study shows that parameters derived from the DTI sequence (MD and FA) do not provide additional information over parameters derived from a multiple b-value sequence (from two-and three-component fits). Hence, the DTI sequence can be omitted, decreasing total scanner time with one third, to about 30 min.

Conclusion
In conclusion, parameters derived from a two-or three-component fit of the diffusion signal are sensitive to histopathological features of kidney lesions.

Additional files
Additional file 1: Figure S1. Diffusion-derived parameter maps of each tumor type: an unaffected kidney (A-D), RCC (E-I), clear cell renal cell carcinoma (cc-RCC) with sarcomatoid differentiation (J-N), papillary cell clear cell carcinoma (O-S), cc-RCC with papillary growth (T-X), hemangioma (Y-Ba), simple cyst (Ca-Fa), RCC with micro cysts (Ga-Ka), oncocytoma (La-Pa). First row: gross appearance of the whole kidney or tumor after nephrectomy, second row: anatomical reference (after processing) which is used to manually draw a mask of the whole kidney and tumor (T 2 -TSE), third row: the unweighted image of the diffusion scan after processing and masking (DWI-b0), fourth row: a merge of the fraction maps from the tri-exponential fit, red = f fast, blue = f interm, green = f slow (1-f interm -f fast ), fifth row: diffusion coefficient from the tri-exponential fit (D tri ). (TIF 36317 kb) Additional file 2: Figure S2. D mono plotted against other diffusion coefficients MD (A), D bi (B) and D mono (C) for each lesions. MD versus D mono displays good correlation, whereas D bi and D tri are structurally lower. Cor-heal = healthy cortex, med-heal = healthy medulla, ccRCC = clear cell renal cell carcinoma, pRCC = papillary cell Rhema = haemangioma, onco = oncocytoma, RCC, ccRCC-pRCC = ccRCC with papillary growth, ccRCC-sarc = ccRCC with sarcomatoid differentiation, ccRCC-cyst is ccRCC with micro-cystic structures. (TIF 1288 kb) Additional file 3: Table S1.