Perfusion MR imaging detection of carcinoma arising from preexisting salivary gland pleomorphic adenoma by computer-assisted analysis of time-signal intensity maps

Tumor perfusion can be evaluated by analyzing the time-signal intensity curve (TIC) after dynamic contrast-enhanced (DCE) MR imaging. Accordingly, TIC profiles are characteristic of some benign and malignant salivary gland tumors. A carcinoma ex pleomorphic adenoma (CXPA) arises from a long-standing pleomorphic adenoma (PA) and has a distinctive prognostic risk depending on the tumor growth potential such as invasion beyond the preexisting capsule. Differentiating CXPA from PA can be very challenging. In this study, we have attempted to discriminate CXPA from PA based on a two-dimensional TIC mapping algorithm. TIC mapping analysis was performed on 8 patients with CXPA and 20 patients with PA after dynamic contrast-enhanced (DCE) MR imaging using a 1.5-T MR system. The TIC profiles obtained were automatically categorized into 5 types based on the enhancement ratio, maximum time, and washout ratio (Type 1 TIC with flat profile, Type 2 TIC with slow uptake, Type 3 TIC with rapid uptake and a low washout ratio, Type 4 TIC with rapid uptake and a high washout ratio, and Type 5 TIC not otherwise specific). The percentage tumor areas with each of the 5 TIC types were compared between CXPAs and PAs. Stepwise differentiation and cluster analysis using multiple TIC cut-off thresholds distinguished CXPAs from PAs with 75% sensitivity, 95% specificity, 86% accuracy, and 86% positive and 90% negative predictive values, when tumors with ≤1.1% Type 1 and ≥15% Type 4, or those with ≤1.1% Type 1, ≥78.1% Type 2, ≥16.1% Type 3, and <15% Type 4, or those with >1.1% Type 1, ≥78.1% Type 2, and ≥16.1% Type 3 areas were diagnosed as CXPAs. The overall TIC profiles predicted some aggressive CXPA growth patterns. These results suggest that stepwise differentiation based on TIC mapping is helpful in differentiating CXPAs from PAs.


MR imaging
MR imaging was performed using a 1.5-T MR imager (Gyroscan Intera 1.5T Master, Philips Healthcare, Best, The Netherlands) with a 140 mm × 170 mm Synergy Flex M coil (Philips Healthcare). T1-and fat-suppressed (spectral attenuated with inversion recovery, SPAIR) T2-weighted MR images (TR/TE/number of signal acquisition = 500 ms/15 ms/2, and 6385 ms/80 ms/2, respectively) were obtained by using a turbo spin-echo (TSE) sequence (TSE factor = 3 and 15, respectively). We used a 200-mm FOV, 256 × 204 acquisition and 512 × 512 reconstruction matrix sizes, a 4-mm slice thickness and a 0.4-mm slice gap. For DCE studies, axial T1-weighted images (TR/TE/number of signal acquisitions = 306 ms/10 ms/1) were obtained using a TSE sequence (TSE factor = 7), 200-mm field of view (FOV), 4-to 5-mm slice thickness, 0.4-to 0.5-mm slice gap, and a 256 × 180 matrix size. Gadolinium (gadopentetate dimeglumine, Magnevist, Bayer HealthCare) was injected intravenously at a dose of 0.2 ml/kg body weight and at an injection rate of 1.5 ml/s, followed by a 20-ml saline flush. In each patient, 19 MR data-acquisitions were obtained at an interval of 10 s (0~180 s). We analyzed 3 slices from the mid-portion of each tumor. The MR images used in the study were assessed from the Nagasaki University Hospital database and the patient data were anonymised.

Conventional MR image characterization
T1-and fat-suppressed T2-weighted MR images of CXPAs and PAs were evaluated by 2 radiologists with 17 and 22 years of experience in the field of head and neck radiology. Consensus reading of the images was based on the signal properties of tumor parenchyma (homogeneous or heterogeneous relative to the neighboring muscle on fat-suppressed T2-weighted MR images) and on the signal characteristics of the tumor capsule (the presence or absence of contrast-enhanced tumor demarcation surrounding the entire tumor perimeter at the maximum tumor area on CE T1-weighted MR images).

Time signal intensity curve (TIC) analysis
The sequential dynamic MR images in a Digital Imaging and Communications in Medicine (DICOM) format were transferred onto a personal computer and were then analyzed by using the ImageJ (NIH, https://imagej.nih.gov/ij/) and Mathematica (Wolfram Research) softwares. In this analysis system, an ROI was manually placed onto a tumor area to include as much tumor area as possible (not excluding necrotic/cystic parts) on a pre-enhanced (= 0 s) T1-weighted MR image, and it was automatically placed onto the subsequent 18 dynamic MR images by repeated copying and pasting the initial ROI using the ImageJ software. Subsequently, the tumor areas were extracted using the ImageJ software. The TIC analysis was performed by using the Mathematica software on a pixel-by-pixel basis (TIC mapping) as follows.
Each of the obtained TICs were automatically classified into 5 types (Types 1-5) based on the enhancement ratio (ER), maximum time (MT), washout ratio (WR) (Fig 1). These TIC parameters were calculated using the following equations: ER (%) = (maximum signal intensity -pre-enhancement signal intensity) × 100/signal intensity of pre-enhancement; MT = time corresponding to the maximum signal intensity; WR (%) = (maximum signal intensity -signal intensity 180s after the start of contrast medium injection) × 100 /(maximum signal intensity -pre-enhancement signal intensity).
Type 1 TICs are those with ERs smaller than or equal to 20%. Type 2 TICs are those with ERs greater than 20% and MTs equal to or longer than 120 s, Type 3 TICs are those with ERs greater than 20%, MTs shorter than 120 s, and WRs smaller than or equal to 30%. Type 4 TICs are those with ERs greater than 20%, MTs shorter than 120 s, and WRs greater than 30% and smaller than 70%. Type 5 TICs are those with ERs greater than 20%, MTs shorter than 120 s, and WRs equal to or greater than 70% (Fig 1). Accordingly, the TIC patterns were referred to flat (Type 1), slow uptake (Type 2), rapid uptake with a low WR (Type 3), rapid uptake with a high WR (Type 4), or not otherwise specified (Type 5).
The percentage areas corresponding to each of the 5 TIC patterns of the total tumor area (a sum of 3 sequential tumor areas including the maximum tumor area) were calculated, and the two-dimensional distributions of the distinct TIC patterns in the 3 sequential tumor areas were displayed as color TIC maps.
Previous studies have shown that TIC profiles correlate well with the histological components of the tissues [10]. Accordingly, we tentatively categorized the overall TIC profiles of the CXPAs and PAs into 4 types, referring to them as matrix-rich, cell-rich, intermediate, and cystic types. The matrix-rich type indicated tumors with !60% Type 2 areas, the cell-rich type had !15% Type 4 TIC areas, the intermediate type had <60% Type 2 and <15% Type 4 areas, and the cystic type contained !50% Type 1 areas.

Histological assessment of CXPAs
Histologic characteristics of 8 CXPAs were analyzed using hematoxylin-and eosin-stained paraffin-embedded tissues from the mid-portion of the excised tumors. These were based on the presence (invasive) or absence (non-invasive) of extensive tumor invasion (!1.5 cm cancer extension beyond the tumor capsule), histological subtypes of the malignant components (adenocarcinoma, adenoid cystic carcinoma, undifferentiated carcinoma, salivary duct carcinoma, etc.), and the proportion of malignant components within the tumors expressed as a percentage of the total tumor area [11]. Subsequently, we correlated the histological characteristics with the overall TIC profiles of the tumors. TIC profiles were also compared between the preexisting PA and malignant tumor areas by analyzing the TIC map areas that roughly corresponded to either of the tumor areas within the 8 CXPAs.

Statistical analysis
Statistical analysis was performed using the JMP software (version 11, SAS, Chicago, IL). TIC cut-off thresholds for differentiating between CXPAs and PAs were determined using receiver operating characteristic (ROC) analysis, followed by cluster analysis to perform a stepwise differentiation of CXPA from PA using multiple TIC criteria. Average percent tumor areas with distinctive TIC patterns of CXPA and PA were compared using the Welch's t-test. The sensitivity, specificity, accuracy, and positive and negative predictive values for the single or multiple cut-off thresholds to diagnose CXPA or PA were studied. For all comparisons, a value of p<0.05 was considered statistically significant.

Conventional MR imaging findings of CXPAs and PAs
Findings in terms of the presence or absence of a contrast-enhanced capsule that surrounds the entire tumor perimeter, and of the homogeneous or heterogeneous tumor signal characteristics are summarized in Table 1. We found that both the MR imaging criteria were not helpful for differentiating between the two tumors (Fig 2A-2C, 2F-2H and 2K-2M).

Diagnostic performance of single TIC criteria
Detailed TIC profiles of the 8 patients with CXPA and 20 patients with PA are shown in Table 2. The average percent tumor areas with distinctive TIC patterns (Type 1-Type 5) were not significantly different between CXPAs and PAs. ROC analysis indicated that 1.1% Type  (Table 3 and Fig 2D, 2I and 2N). However, the single TIC cut-off thresholds provided low or at best moderate diagnostic abilities for differentiating between the two salivary gland tumor entities.

Diagnostic performance of multiple TIC criteria in stepwise approach
We next tested whether combinations of the TIC indices determined by the ROC analysis could improve the diagnostic abilities for differentiating between CXPAs and PAs. To this end, we used a stepwise differentiation strategy with different combinations of TIC cut-off thresholds. Cluster analysis indicated the dendrogram that most effectively differentiated CXPAs from PAs (Fig 3). The stepwise differentiation strategy could differentiate CXPAs from PAs with 75% sensitivity, 95% specificity, 86% accuracy, and 86% positive and 90% negative predictive values, when tumors with 1.1% Type 1 and !15% Type 4, or those with 1.1% Type 1, !78.1% Type 2, !16.1% Type 3, and <15% Type 4, or those with >1.1% Type 1, !78.1% Type 2, and !16.1% Type 3 areas were diagnosed as CXPAs. (Table 3). However, 3 CXPAs and 1 PA were still misdiagnosed using this strategy (Fig 3).

Correlations of overall TIC profiles and histological features of CXPA
When the tumors were categorized into any of the 3 overall TIC types (matrix-rich, cell-rich, and intermediate types), those with intermediate overall TIC types had the most invasive histological characteristics, such as invasive tumor margins and greater proportions of malignant components within the tumor (Table 4, Fig 2E, 2J and 2O). However, the small study population of CXPA did not allow us to reach a conclusion in terms of statistically significant correlations between the aggressive histological characteristics and the overall TIC profiles.
The progression from PA to CXPA is usually due to the emergence of an epithelial, malignant component within a pre-existing, usually highly hyalinized PA. The TIC patterns are based in the proportion of extracellular matrix versus cellular compartment. Therefore, we reasoned that the malignant progression is associated with TIC profile transition from Type 2 to Type 3 or 4 patterns. We found that the pre-existing PA areas exhibited Type 2 overall TIC profiles, while most of the malignant areas showed Type 3 overall TIC profiles or Type 2 overall TIC profiles with larger Type 3/4 TIC tumor areas compared with the corresponding preexisting PA areas (Fig 4).

Discussion
In the present study, we aimed to differentiate CXPA from PA using pixel-based TIC analysis. The stepwise approach with multiple TIC cut-off thresholds allowed us to differentiate CXPAs from PAs with 75% sensitivity, 95% specificity, 86% accuracy, 86% positive and 90% negative predictive values. The obtained diagnostic ability was higher than that with a single use of TIC thresholds. However, the diagnostic accuracy remained less than 90% and 3 out of the 8 CXPA cases were misdiagnosed using the stepwise approach.
The TIC analysis has been most frequently used for assessing DCE MR images of salivary gland tumors. In a TIC, the early slope phase corresponds to the contrast agent transfer to the tissues, and it is considered to be largely dependent upon the blood flow rate, inflow blood volume, and capillary permeability [8]. The magnitude of the maximum signal intensity is representative of the volume of the extravascular-extracellular space, whereas the maximum time is related to the extravascular-extracellular space per contrast agent transfer to the tissues.  Table 2). Boxes directly right to the respective photomicrographs (H & E staining) show overall TIC profiles (upper panel in each box) and TIC mapping (lower panel in each box; % tumor areas as Type 1/Type 2/Type 3/Type 4/Type 5) of pre-existing PA (green, PA) and malignant tumor (blue, C) areas within CXPA tumors. https://doi.org/10.1371/journal.pone.0178002.g004 Perfusion MR imaging of salivary gland tumors based on time-signal intensity curve mapping Although the capillary density and blood volume of the tumors are closely related to the kinetics of contrast enhancement, cellularity can also contribute to the early slope phase. In cancers, for example, high capillary permeability and poor lymphatic drainage create a condition of limited cellularity [12]. Low cellularity may lead to slow uptake of the contrast agent. Meanwhile, the washout ratio depends upon the histological type, such as phenotypes of the cancer cells and properties of the cancer stroma. Therefore, the mechanisms that influence the washout ratio of the contrast agent in these tumors are complex. Contrast enhancement patterns can also be greatly influenced by the vascular permeability of the lesions. Capillary density may be a minor physiologic factor compared to vascular permeability with regard to the differences in TIC profiles [13,14].
In a pixel-based TIC analysis, cystic/necrotic areas display Type 1 TIC patterns and cancer nests containing loosely packed cancer cells associated with extensive fibrotic/myxoid tissues may have Type 2 or Type 3 TIC patterns depending on the density of cancer nests and fibrosis. Tumor areas with densely packed cancer cells are often associated with Type 4 TIC patterns [15,16]. The cancer focus area in CXPA varies greatly from one tumor to another. In addition, PAs also have histologic features with a wide variety of benign components [4]. Therefore, a single TIC criterion is not sufficient for the effective differentiation of CXPA from PA. The stepwise approach with multiple TIC indices permitted effective differentiation of CXPAs from PAs. The overall profiles of CXPAs can be categorized into either of the 3 types with varying combinations of tumor areas of distinctive TIC profiles; these included (a) cell-rich type with !15.0% Type 4 areas, (b) matrix-rich type with !60% Type 2 areas, and (c) intermediate type with <60% Type 2 and <15% Type 4 areas. However, PAs had an additional TIC type with !50% Type 1 areas (cystic type). The present study findings implied that the intermediate type with rich cancer nests and matrix might represent a more aggressive form of CXPA. This is consistent with the previous report that aggressively growing cancer foci in metastatic nodes exhibited Type 2 and Type 3-dominant TIC profiles [15]. Furthermore, comparison of TIC profiles between pre-existing PAs and malignant tumor areas within CXPAs in the present study supports the notion that malignant transformation from the pre-existing PA is associated with a transition of the tumor perfusion characteristics (from Type 2 to Type 3/4 TIC profiles). However, any causal relationship was found between the false negative results of the 3 patients (no. 2, 5, and 7) and the TIC profiles of the pre-existing PA and malignant tumor areas (see Table 2 and Fig 4B, 4E and 4G).
The fact that 3 cases out of the 8 CXPA patients were misdiagnosed using the stepwise differentiation may be a very critical problem for the preoperative imaging diagnosis. In addition, the false-negative CXPA cases had no consistent TIC or histological characteristics in the present study. The poor prognostic factors for CXPA reportedly included extracapsular extension of cancer components >2.5 mm from the capsule of the preexisting PA [6]. However, the detection of extracapsular extension using MR imaging can be very challenging; in fact, the morphology of the tumor margin and capsule was indistinguishable between CXPA and PA. In the present study, we used a higher threshold of 1.5 cm for the invasive nature of the tumor [11], and found that the overall TIC profiles moderately correlated with the tumor encapsulation characteristics. Another prognostic factor is the presence of myoepithelial carcinoma in CXPA [17]. The clinical behavior of minimally or non-invasive CXPA is not aggressive compared to that of widely invasive CXPA [5,18]. However, the presence of myoepithelial carcinoma subtype within a minimally invasive CXPA may increase the risk of recurrence [17]. In contrast, Lewis et al. denied the prognostic importance of histological subtypes in malignant components [11]. In the present study, we did not find any obvious correlations between the histological subtypes of malignant components and the TIC profiles. However, another study with a large cohort of CXPA patients is needed to confirm this finding.
A major limitation of this study is that the diagnostic ability achieved via the stepwise differentiation with multiple TIC indices was not sufficiently high, and 3 out of 8 CXPAs were misdiagnosed using this approach. Related to this shortcomings, CXPA is an infrequent tumor and higher numbers are difficult to achieve. The subdivision into further categories, both of TIC patterns and histological features leads to very low numbers in each category. Therefore, small numbers of false positive/negative cases greatly affect the results. Water diffusion is the physiological property that can be predicted separately from tissue perfusion on diffusionweighted MR imaging depending upon the b-values used [19]. Incorporation of the tumor diffusion characteristics in the criteria could achieve further improvement in the diagnostic ability for differentiating between CXPAs and PAs. The diffusion coefficient of salivary gland tumors used as as a single criterion or combined with TIC characteristics can effectively distinguish between benign and malignant tumors [16,20,21]. Furthermore, diffusion coefficient values were found to correlate well with the histological types of salivary gland tumors [10]. This hypothesis should also be evaluated further in a future study.
The multi-step differentiation system using different TIC patterns may be boring for physicians and radiologists, and different TIC cut-off thresholds may be needed for a different patient cohort. However, the present study has confirmed the usefulness of the computerassisted detection/differentiation algorithms (TIC and cluster analyses) in differentiating CXPA from PA that had complex and similar perfusion MR imaging characteristics. These results suggested that the machine learning algorithm would be promising for the effective diagnosis of salivary gland tumors with incredibly complex and multi-dimensional medical imaging patterns that exceed human pattern recognition capabilities [22].

Conclusions
In this study, we showed that stepwise differentiation using multiple TIC cut-off thresholds differentiated CXPA from PA with 75% sensitivity, 95% specificity, 86% accuracy, and 86% positive and 90% negative predictive values. It should be noted that the obtained diagnostic ability cannot be guaranteed in a different patient cohort. However, at present we have no imaging tools that has been reported to be promising for the preoperative diagnosis of CXPA in a patient with long-standing PA, other than MR imaging. Estimating the perfusion property may be effective for predicting CXPA and be useful for planning the management of patients with long-standing PA.