Diagnostic Performance of Diffusion Kurtosis Imaging for Benign and Malignant Breast Lesions: A Systematic Review and Meta-Analysis.

Purpose: Magnetic resonance imaging (MRI) has a high sensitivity for differentiating between malignant and non-malignant breast lesions but is sometimes limited due to its low specificity. Here, we performed a meta-analysis to evaluate the diagnostic performance of mean kurtosis (MK) and mean diffusivity (MD) values in magnetic resonance diffusion kurtosis imaging (DKI) for benign and malignant breast lesions.
Methods: Original articles on relevant topics, published from 2010 to 2019, in PubMed, EMBASE, and WanFang databases were systematically reviewed. According to the purpose of the study and the characteristics of DKI reported, the diagnostic performances of MK and MD were evaluated, and meta-regression was conducted to explore the source of heterogeneity.
Results: Fourteen studies involving 1,099 (451 benign and 648 malignant) lesions were analyzed. The pooled sensitivity, pooled specificity, positive likelihood ratio, and negative likelihood ratio for MD were 0.84 (95% confidence interval (CI), 0.81-0.87), 0.83 (95% CI, 0.79-0.86), 4.44 (95% CI, 3.54-5.57), and 0.18 (95% CI, 0.13-0.26), while those for MK were 0.89 (95% CI, 0.86-0.91), 0.86 (95% CI, 0.82-0.89), 5.72 (95% CI, 4.26-7.69), and 0.13 (95% CI, 0.09-0.19), respectively. The overall area under the curve (AUC) was 0.91 for MD and 0.95 for MK. Conclusions: Analysis of the data from 14 studies showed that MK had a higher pooled sensitivity, pooled specificity, and diagnostic performance for differentiating between breast lesions, compared with MD.

1. Introduction

Magnetic resonance imaging (MRI) of the breast has been shown to have high sensitivity but relatively low specificity for characterizing breast lesions, which urges the need for additional imaging techniques or biopsy [1]. Since the past decades, a lot of efforts have been directed toward developing techniques for improving the specificity of breast MRI to differentiate between malignant and nonmalignant breast lesions [2-4].

The use of apparent diffusion coefficient (ADC) values, calculated by fitting diffusion-weighted imaging signals in a single exponential model, has improved the diagnostic accuracy for differentiating between breast lesions [5-8]. However, the standard ADC is based on free diffusion under ideal conditions [9, 10]. In biological tissues, diffusion is limited by various barriers and presents a non-Gaussian distribution. Therefore, in some cases, ADC may not accurately reflect the microstructure of biological tissues and more advanced magnetic resonance diffusion imaging techniques are needed [9, 10].

Recently, there has been an increasing interest in diffusion kurtosis imaging (DKI) to assess non-Gaussian diffusion behavior in diseases involving complex biological tissues such as breast cancer [11-13]. DKI technology has been shown to improve the sensitivity and accuracy for diagnosing benign and malignant breast lesions and is therefore regarded as a promising technique for the differential diagnosis of breast lesions [14, 15]. DKI can measure molecular water movement via ADC values. The monoexponential model used in conventional DWI assumed that the microenvironment is homogeneous and that the diffusion of water molecules follows a Gaussian distribution, which causes a linear decay of the logarithm of the DWI signal intensity as the b value increases.

DKI can be used to describe non-Gaussian distributions, using parameters such as mean kurtosis (MK) and mean diffusivity (MD) [16, 17]. In DKI, the MD value of the diffusion coefficient represents the ADC value, and when corrected by non-Gaussian distribution, it can reflect the overall diffusion level and diffusion resistance of water molecules. MD represents the average diffusion coefficient, and a decrease in its diffusion level increases the diffusion resistance, which lowers the MD value. MK is used to evaluate the degree of deviation of diffusion displacement distribution of water molecules based on a Gaussian function. It mainly reflects microstructure complexity. When the region of interest (ROI) is more complex, the diffusion of water molecules is more limited, and the MK value increases [18-21]. The European Society of Breast Radiology working group has confirmed the importance of breast using DWI in breast MRI protocol to help differentiate between different types of breast lesions, distinguishing in situ from invasive lesions, and even predicting the responses to neoadjuvant therapy [22].

In previous studies, the application of DKI for identifying breast lesions has been well-reported, but information related to variation in diagnostic values of DKI and the advantages and disadvantages have not yet been comprehensively evaluated. Therefore, we conducted a review of the recent literature and meta-analysis to compare the values of MD and MK in differentiating benign and malignant breast lesions.

2. Materials and Methods

2.1. Methodology of Searching the Literature

We performed a systematic review of original articles from PubMed, EMBASE, and WanFang databases that were published from 2010 to 2019. No relevant literature was found before 2010. The following keywords were used: “DKI,” “Diffusion Kurtosis Imaging,” “breast,” and “non-Gaussian distribution.” All keywords listed in the document search were used in combination. The search was restricted to studies in English and Chinese and studies with subjects limited to “humans.” All articles were independently evaluated by two radiologists with 10 and 12 years of experience, respectively. The full texts of the retrieved articles deemed eligible were retrieved. References of the retrieved papers were hand-searched for additional eligible studies. The study was conducted in accordance with PRISMA guidelines.

2.2. Eligibility Criteria

Articles were selected based on the following inclusion criteria: (1) evaluated the diagnostic performance of DKI for benign and malignant breast lesions; (2) included the common pathological types of benign and malignant breast lesions and pathological results that were used for diagnosis; (3) included more than 20 cases; (4) were categorized as A-grade or B-grade literature (including prospective and retrospective studies); (5) provided quantitative measurements (MD and MK); and (6) for which the true-positive (TP), false-positive (FP), true-negative (TN), and false-negative (FN) parameters of DKI for the diagnosis of benign and malignant breast lesions could be directly or indirectly obtained. We excluded articles published as “reviews,” “letters,” “comments,” “editorials,” or “case reports.”

2.3. Data Extraction and Quality Appraisal

Extracted information included the study author(s), publication date, country, number of cases, average age and age range of the subjects, continuity and type of the study (prospective or retrospective), MRI machine type, and MRI equipment manufacturer. Diagnostic performance parameter information extraction included TP, FP, TN, FN, sensitivity, and specificity. The included studies were evaluated following the Quality Assessment of studies of Diagnostic Accuracy included in the Systematic Reviews (QUADAS-2 tool) checklist using the Review Manager 5.3 software (Northern Europe, The Cochrane Center).

2.4. Statistical Analysis

Heterogeneity between the trials was analyzed using the chi-squared and I2 tests. Results showed that P > 0.1 and I2 < 50% indicated no significant heterogeneity and P ≤ 0.1 and I2 > 50% indicated substantial heterogeneity. The sources of heterogeneity were determined using metaregression analysis, with the results expressed as relative diagnostic performance odds ratios.

The Meta-DiSc 14.0 software (Unit of Clinical Biostatistics, Ramón y Cajal Hospital, Madrid, Spain) was used to pool the diagnostic variables, including pooled sensitivity, pooled specificity, positive likelihood ratio (LR), and negative LR. The results are presented as forest plots, which show the results of individual studies with the corresponding 95% confidence intervals (CI). Generally, a positive LR > 5.0 and a negative LR < 0.2 were considered clinically significant. A higher sensitivity and specificity indicated good diagnostic performance.

We also calculated the area under the curve (AUC) values, which were graded as follows: 0.9-1.0, excellent; 0.8-0.9, good; 0.7-0.8, fair; 0.6-0.7, poor; and 0.5-0.6, useless. A Fagan plot was created for post-test probability assessment using the Stata 11.2 software (Stata Corp., College Station, TX, USA). Publication bias was assessed using Deek's funnel plot. P values < 0.05 were considered statistically significant.

3. Result

3.1. Literature Search

In the initial literature search, 412 articles were identified, of which 383 articles, reviews, and abstracts were excluded. Of the remaining 29 related articles, 15 were excluded due to incomplete data or unqualified cases. Finally, 14 articles [19–21, 23–33] were found eligible for this study, of which seven were written in English and seven in Chinese. The literature-screening process and results are shown in Figure 1.

Figure 1

Flow diagram of study selection for meta-analysis.

3.2. Characteristics of Analyzed Studies

The characteristics of the 14 studies are summarized in Table 1. Of them, six were retrospective and eight were prospective studies. In all, the data of 1,023 patients were included in this meta-analysis and 1,099 breast lesions were assessed by DKI, of which 648 (58.9%) were cancerous lesions. All the included studies used surgical pathology as the standard reference for diagnosing breast lesions. Eleven studies assessed the diagnostic accuracy of DKI using the more sensitive 3.0 Tesla MRI system, while there were three studies that used the 1.5 Tesla MRI system.

Table 1

Characteristics of the included studies in this meta-analysis.

Included studies	Year	Country	Age (years)	No. of cases	MRI apparatus, field strength b value (s/mm2)	Histological analysis time	Study design
Liu et al.	2017	China	43 (18-69)	37	Siemens 1.5 T0, 700, 1400, 2100, 2800	PostoperativeBiopsy	Prospective
Chen et al.	2016	China	46 (18-82)	45	GE 3.0 T0, 600, 1200, 1800	PostoperativeBiopsy	Prospective
Wang et al.	2017	China	42 (18-60)	45	GE 3.0 T0, 500, 1000, 1500, 2000, 2500	PostoperativeBiopsy	Retrospective
Lin et al.	2017	China	45 (21-69)	53	GE 3.0 T0, 500, 1 000, 2 000, 2500	PostoperativeBiopsy	Retrospective
Gao et al.	2017	China	42 (13-61)	72	Philips 3.0 T0, 500, 800, 2000	PostoperativeBiopsy	Retrospective
Li et al.	2018	China	46 (22-76)	64	Philips 3.0 T0, 500, 1000, 1500, 2000, 2500	PostoperativeBiopsy	Retrospective
Li et al.	2016	China	N	137	Siemens 3.0 T0, 500, 800, 1000, 2000	PostoperativeBiopsy	Prospective
Nogueira et al.	2014	Porto	N	36	Siemens 3.0 T600, 800, 1000, 2000, 3000	PostoperativeBiopsy	Prospective
Sun et al.	2015	China	45 (19-70)	98	Siemens 1.5 T0, 700, 1400, 2100, 2800	PostoperativeBiopsy	Retrospective
Wu et al.	2014	China	57 ± 14	103	Siemens 3.0 T0, 500, 750, 1000, 1500, 2000	PostoperativeBiopsy	Prospective
Li et al.	2018	China	22–79	106	Philips 3.0 T0, 500, 1000, 1500, 2000, 2500	PostoperativeBiopsy	Prospective
Christou et al.	2017	UK	37-71	49	1.5 T MRI0, 400, 800, 1100, 1300	PostoperativeBiopsy	Prospective
Liu et al.	2019	China	13–64	71	3.0 T MRI0, 500, 800, 2000	PostoperativeBiopsy	Retrospective
Li et al.	2019	China	N	120	3.0 T MRI0, 500, 1000, 1500, 2000, 2500, 3000	PostoperativeBiopsy	Prospective

N: not reported.

3.3. Quality of the Included Studies

The QUADAS-2 scale shows that the included studies were of acceptable methodological quality. Figure 2 shows the QUADAS–2 results regarding the proportion of studies with low, high, or unclear risk of bias and applicability concerns. The results showed a high risk of bias in patient selection and reference standards, and a high applicability concern existed in the reference standard.

Figure 2

Quality assessment of the included studies using the QUADAS-2 tool.

3.4. Threshold Effect and Heterogeneity Assessment

Spearman's rank correlation coefficient between the logit of sensitivity and the logit of “1–specificity” for MD was -0.244 (P = 0.400). There was no threshold effect. Heterogeneity between the included studies was investigated, and we observed that while the pooled specificity (Q = 15.06, P = 0.304, I2 = 13.7%) and positive LR (Q = 15.11, P = 0.300, I2 = 14.0%) of the studies were homogenous, there was some heterogeneity in their pooled sensitivity (Q = 47.59, P ≤ 0.001, I2 = 72.7%) and negative LR (Q = 44.09, P < 0.001, I2 = 70.5%).

Spearman's rank correlation coefficient between the logit of sensitivity and the logit of “1–specificity” for MK was -0.113 (P = 0.699), and no threshold effect was detected. The heterogeneity between the included studies was assessed, and similar to MD, the pooled specificity (Q = 17.17, P = 0.192, I2 = 24.3%) and positive LR (Q = 19.80, P = 0.100, I2 = 34.3%) were homogenous, while some heterogeneity was observed in the pooled sensitivity (Q = 35.00, P = 0.001, I2 = 62.9%) and negative LR (Q = 33.71, P = 0.001, I2 = 61.4%).

3.5. Subgroup Analysis

Based on the research features, metaregression, subgroup, and regression analyses of MD and MK values were performed for each study type (prospective, retrospective), number of lesions (≥60, ≤60), field strength (3.0 T, 1.5 T), and b values (0-1500, 0-2000, 0-2500, and 0-3000 s/mm2). No statistical significance was observed, suggesting that these factors were not the source of heterogeneity (Table 2).

Table 2

Assessment of the diagnostic accuracy and heterogeneity in subgroup analysis.

Aggregate effect	Cases	MD			MK
		PooledSensitivity(95% CI)	PooledSpecificity(95% CI)	P value	PooledSensitivity (95% CI)	PooledSpecificity(95% CI)	P value
Study type				0.83			0.47
Prospective	8	0.87 (0.83-0.91)	0.81 (0.74-0.87)		0.92 (0.88-0.95)	0.86 (0.80-0.91)
Retrospective	6	0.87 (0.81-0.92)	0.84 (0.76-0.89)		0.89 (0.83-0.93)	0.91 (0.84-0.95)
Number of lesions(a)				0.21			0.27
≥60	8	0.87 (0.83-0.91)	0.81 (0.75-0.85)		0.90 (0.87-0.93)	0.87 (0.83-0.91)
≤60	6	0.87 (0.79-0.93)	0.91 (0.78-0.97)		0.93 (0.86-0.97)	0.93 (0.81-0.98)
Field intensity				0.12			0.13
3.0 T	11	0.93 (0.87-0.97)	0.88 (0.78-0.95)		0.95 (0.89-0.98)	0.93 (0.84-0.97)
1.5 T	3	0.85 (0.81-0.88)	0.80 (0.74-0.85)		0.89 (0.86-0.93)	0.87 (0.82-0.91)
b values (s/mm2)				0.65			0.43
0-1500	1
0-2000	5	0.87 (0.82-0.91)	0.83 (0.75-0.89)		0.92 (0.88-0.95)	0.88 (0.81-0.94)
0-2500	5	0.90 (0.84-0.95)	0.88 (0.77-0.95)		0.93 (0.87-0.97)	0.92 (0.83-0.98)
0-3000	3	0.81 (0.70-0.89)	0.74 (0.64-0.83)		0.82 (0.71-0.89)	0.84 (0.75-0.91)

3.6. Overall Diagnostic Performance

Quantitative data for DKI from MRIs were retrieved to assess the diagnostic performances of MD and MK in differentiating between malignant and benign breast lesions. The pooled sensitivity, specificity, positive LR, negative LR, and summary receiver operating characteristic (SROC) curve for diagnosing malignant lesions were calculated using TF, FP, TN, and FN (Table 3).

Table 3

Diagnostic performance of MD and MK for the included studies.

Included studies	MD						MK
	TP	FP	FN	TN	Sensitivity	Specificity	TP	FP	FN	TN	Sensitivity	Specificity
Liu et al. 2017	25	2	3	7	0.89	0.79	26	1	2	8	0.93	0.89
Chen et al. 2016	14	5	3	25	0.83	0.85	15	5	2	25	0.92	0.84
Wang et al. 2017	19	4	3	27	0.86	0.87	21	11	1	20	0.95	0.65
Lin et al. 2017	31	1	7	14	0.82	0.93	34	1	4	14	0.89	0.93
Gao et al. 2017	36	6	7	26	0.84	0.81	36	2	7	30	0.84	0.94
Li et al. 2018	24	8	6	26	0.80	0.77	25	5	5	29	0.83	0.85
Li et al. 2016	108	6	7	32	0.94	0.84	110	7	5	31	0.96	0.82
Nogueira et al. 2014	29	2	2	11	0.93	0.85	28	2	3	11	0.91	0.89
Sun et al. 2015	55	5	2	36	0.97	0.88	54	3	3	38	0.95	0.93
Wu et al. 2014	65	7	17	35	0.79	0.84	75	4	7	38	0.92	0.92
Li et al. 2018	38	16	9	43	0.81	0.73	38	10	9	49	0.81	0.83
Christou et al. 2017	31	1	3	18	0.91	0.94	33	1	1	18	0.97	0.94
Liu et al. 2019	35	9	7	21	0.83	0.70	35	5	7	25	0.83	0.83
Li et al. 2019	37	6	25	52	0.60	0.90	44	8	18	50	0.71	0.86

MK: mean kurtosis; MD: mean diffusion; TP: true-positive; FP: false-positive; TN: true-negative; FN: false-negative.

Analysis of the 14 studies showed that their pooled sensitivity, pooled specificity, positive LR, and negative LR for MD were 0.84 (95% CI, 0.81-0.87), 0.83 (95% CI, 0.79-0.86), 4.44 (95% CI, 3.54-5.57), and 0.18 (95% CI, 0.13-0.26), respectively (Figure 3). Using the ROC curve, the overall AUC was calculated as 0.91 (standard error = 0.01) (Figure 4(a)). For MK, the pooled sensitivity, pooled specificity, positive LR, and negative LR were 0.89 (95% CI, 0.86-0.91), 0.86 (95% CI, 0.82-0.89), 5.72 (95% CI, 4.26-7.69), and 0.13 (95% CI, 0.09-0.19), respectively (Figure 5). Using the SROC curve, the overall AUC was calculated as 0.95 (standard deviation = 0.01) (Figure 6(a)).

Figure 3

Forest plots of pooled sensitivity, pooled specificity, positive LR, and negative LR for MD in detecting benign and malignant breast lesions. LR: likelihood ratio; MD: mean diffusivity.

Figure 4

The SROC curve, publication bias, and Fagan Nomogram for MD in detecting benign and malignant breast lesions. SROC: summary receiver operating characteristic; MD: mean diffusivity.

Figure 5

Forest plots of pooled sensitivity, pooled specificity, positive LR, and negative LR for MK in detecting benign and malignant breast lesions. LR: likelihood ratio; MK: mean kurtosis.

Figure 6

The SROC curve, publication bias, and Fagan Nomogram for MK in detecting benign and malignant breast lesions. SROC: summary receiver operating characteristic; MK: mean kurtosis.

Publication bias was assessed using Deek's funnel plot. None of the quantitative analysis results from the 14 studies showed statistical significance for MD or MK (P = 0.82, P = 0.54), suggesting no publication bias between the included studies (Figures 4(b) and 6(b)).

According to the Fagan Nomogram (Figures 4(c) and 6(c)), the possibility of malignant breast lesions detected by conventional MRI was 20%. Our results showed that when the MD value was positive, the malignant probability increased to 56%, and when the MD value was negative, the malignant probability was reduced to 4%. Further, when the MK value was positive, the malignant probability increased to 66%, and when the MK value was negative, the malignant probability decreased to 2%.

4. Discussion

In recent years, there have been many reports on the differential diagnosis of benign and malignant breast lesions using MK and MD values, which have reported improved sensitivity and accuracy in differentiating between benign and malignant breast lesions [1, 2, 14]. Nogueira et al. [21] reported that MK was one of the representative parameters of DKI technology and a useful index to measure the complexity of organizational structure. Li et al. [31] reported that the MK value had the best performance for diagnosing benign and malignant breast lesions. However, Li et al. [29] reported that the MD value had the best performance. Therefore, we conducted this meta-analysis to investigate the clinical reliability of MK and MD in the diagnosis of benign and malignant breast lesions.

Overall, our meta-analysis showed that MD and MK had good sensitivity (84%, 83%) and specificity (89%, 86%) for differentiating between different types of breast lesions. Further, we observed that the sensitivity and specificity of MK were higher than MD, and according to the pooled sensitivity and AUC values, the diagnostic performance of MK was higher. Thus, our findings showed that MK had the best performance, consistent with most previous studies [31, 34].

For MD and MK, there was some heterogeneity in the pooled sensitivity and negative LR. Notably, the subgroup analyses of MD and MK showed sensitivity when using b values of 0-2500 s/mm2. MK can consider the heterogeneity and restriction of diffusion and can also reflect biological tissue complexities [35]. Malignant lesions tend to have higher MK values than benign lesions due to structural heterogeneity, high cell density, interstitial vascular proliferation, and complex tissue structure [11, 20]. A significant increase in sensitivity was observed with 3.0 Tesla MRI system than with the 1.5 Tesla MRI system. In a study by Hur et al. [36], the authors compared the sensitivity of 3.0 Tesla (T) MRI to that of 1.5 T MRI. They found that 3.0 T MRI was more sensitive than 1.5 T MRI and recommended 3.0 T MRI as the preferred imaging modality. The superiority of 3.0 Tesla (T) is related to its higher field strength which improves the signal-to-noise ratio and increases image quality, especially for gadolinium-enhancing lesions [37, 38]. In this present study, the study type, number of lesions, field strength, and b values were not the source of heterogeneity. However, none of the quantitative analysis results from these 14 studies showed publication bias.

The advantage of meta-analysis is that it can overcome some shortcomings such as study design, case source, statistical analysis, and small sample size, and it can improve the credibility of study results [17, 39–41]. A limitation of this study was the lack of relevant literature. In addition, most of the studies were based on results from a single country (China). Compared with MD, MK had higher accuracy in the differential diagnosis of benign and malignant breast lesions. Therefore, based on these findings, MK can be used as a representative parameter of DKI technology.

In conclusion, the results of this meta-analysis showed that MK had a higher pooled sensitivity, pooled specificity, and diagnostic performance for differentiating between breast lesions, compared with MD. Subgroup analysis showed that the sensitivities and specificities of MK and MD were improved when using the 3.0 T MRI system and b values of 0-2500 s/mm2. Thus, based on these observations, we propose the use of MK and 3.0 T MRI system to optimize the diagnostic performance of MRI for differentiating between malignant and benign breast lesions.