Histogram analysis of dynamic contrast-enhanced magnetic resonance imaging to predict extramural venous invasion in rectal cancer

Extramural venous invasion (EMVI) is defined as the presence of tumor cells within blood vessels beyond the muscularis propria of the rectal wall in histology [1], which is an independent risk factor for local and distant recurrence [2, 3] and poorer overall survival of rectal cancer [4]. Traditionally, the gold standard to confirm EMVI status was postoperative pathology. Sometimes the obvious EMVI may be difficult to be identified via routine pathological analysis of the resection specimens because of the total destruction of extramural venous architecture by tumor tissue [5]. In addition, the obliteration and fibrosis of the invaded venous may appear after the neoadjuvant chemoradiotherapy (CRT) as mentioned before [6]. Therefore, these may result in false-negative in histopathology. The inaccurate EMVI assessment may influence the risk stratification and therapeutic decision-making [7]. Given its enormous significance, gaining a noninvasive and explicit diagnosis of EMVI status should be a priority for clinicians developing the individual therapeutic plan.

The magnetic resonance imaging-defined EMVI (mrEMVI) scoring system based on changes in vascular caliber and signal features was the primary imaging modality for specific and accurate assessment of EMVI status [8‐10]. In addition, several studies suggested that mrEMVI has the potential to become an imaging biomarker for predicting histological grade, nodal stage, recurrence risk, and survival after neoadjuvant chemoradiotherapy (CRT) [11‐15]. However, compared with postoperative pathology, the identification of EMVI based on conventional MRI can be inevitably misdiagnosed with a relatively low and wide range (28.2–62%) of sensitivity [14‐16], especially among patients with indeterminate mrEMVI scores of score 2(possibly negative) and score 3 (probably positive) in small vessels (≤ 3 mm) perpendicular to the rectal wall [17]. Furthermore, the inflammation, edema, and fibrosis after neoadjuvant CRT may also increase difficulties in mrEMVI assessment [18, 19].

Recently, kinetic model quantitative parameters of dynamic contrast-enhanced MRI (DCE-MRI) has been frequently applied to describe characteristics of microcirculatory perfusion, and oxygenation in tumor via blood flow, capillary permeability, and permeability surface area in vivo [20‐22]. Some studies have reported associations of DCE-MRI quantitative perfusion parameters with EMVI status in rectal adenocarcinoma [23‐25] and reached the AUCs from 0.680 to o.856.

Histogram Analysis (HA), an emerging method that extracts non-visual imaging information and quantifies pixel intensity patterns within the tumor, indicates homogeneity, heterogeneity, asymmetry, vascularity, and necrosis [19, 23]. Features obtained from HA have been used to distinguish malignant tissues and microvascular invasion in the lung, glioma, colon, breast, and liver [26‐29]. In rectal cancer, the applications of HA involve the prediction of locally advanced rectal carcinoma (stage T3-4 and/or N1-2), response after CRT, and overall survival [19, 30‐32]. To our knowledge, the use of HA based on perfusion parameters of DCE-MRI to estimate EMVI status has not been well established in the literature.

Therefore, we aimed to achieve two goals. First, establish the prediction model of HA features from perfusion parameters based on DCE-MRI to identify EMVI status in rectal cancer patients. Second, further explore the diagnostic performance of this model in high-risk patients with suspicious positive mrEMVI findings (scores 2 and 3).

Preoperative identification of EMVI in rectal cancer assumes a key role in accurate risk stratification and treatment decision-making. Our findings demonstrated the potential of HA features from quantitative perfusion parameters based on DCE-MRI to preoperatively distinguish pathological EMVI status in rectal cancer patients. The K^trans kurtosis and K^trans entropy were identified as independent predictors for the occurrence of pathological EMVI. In addition, we demonstrated the good diagnostic performance of this combined HA model for diagnosing EMVI, particularly in subpopulations with indeterminate mrEMVI scores of 2 and 3.

The prior research elucidated that the presence of EMVI has been identified as an independent risk factor for recurrence risk and poor postoperative survival in rectal cancer [2, 4, 7]. However, consistent with previous findings [14‐16], mrEMVI scores could not well satisfactorily coincide with postoperative pathological EMVI outcomes in this study. Subtle changes in signal features within small extramural vessels with slightly expanded contour and caliber cannot be easily distinguished, especially in rectal cancer lesions with mrEMVI score 2 (possibly negative) and score 3(probably positive) [9, 12].

We demonstrated that the lesions of the EMVI positive-group had significantly higher mean values of K^trans and V_e than the EMVI-negative group. Pathologically, the initial formation of EMVI is dependent on tumor angiogenesis, and then tumor cells invade into microvascular and eventually extend beyond the rectal wall [7, 13, 32]. These developing processes might be reflected by alterations of microcirculatory perfusion in tumor tissue. K^trans is the transfer rate from plasma to extracellular extravascular space (EES), which may correlate to tumor capillary permeability and angiogenesis. Higher K^trans represents greater microcirculation perfusion within tumor tissue [32]. V_e represents the fractional volume of the EES. The proliferation of tumor cells might reduce local permeability surface area, diminish microcirculatory perfusion and lead to microscopic necrosis, thus enlarging EES [23, 33]. Obliteration of the function of cell–cell adhesion molecules might also contribute to elevated V_e value [21, 34]. Yu et al. reported significantly higher K^trans and V_e in the EMVI-positive group than in the EMVI-negative group [25]. Chen et al. discovered that the mrEMVI-positive group had significantly higher V_e than the mrEMVI-negative group, but K^trans showed no significant difference [24]. Wei et al. demonstrated that K^trans was the independent predictor of EMVI in rectal cancer [35]. However, the mean value of K _ep did not show a significant difference between the EMVI-positive and EMVI-negative patients. The K _ep value is equivalent to K^trans / V_e and represents the rate constant of contrast agent escape from the EES into the plasma compartment. It was speculated that K^trans and V_e cannot increase unlimitedly with the increasing degree of vascularization and invasion depth of tumor cells [25]. In addition, with the increase of contrast agent concentration in tumor EES from plasma, the pressure difference between inside and outside microvessels decreases, which may diminish the diffusion rate of contrast agent from EES to plasma in a certain extent, that is, affect the K _ep [22, 24, 25].

Our investigation displayed significantly higher mean values of K^trans and V_e resulted suggestive of pathologic EMVI status. However, these two parameters were not identified as independent predictors of EMVI in rectal cancer through the multivariate logistic regression analysis. It was speculated that these two parameters might be mutually correlated and restricted [24]. Also, the mean values of K^trans and V_e might attenuate differences in perfusion performance between the regional highly vascularized areas of EMVI and other lesions.

Perfusion heterogeneity reflected by HA features obtained on DCE-MRI parameters may provide additional information to improve the diagnostic accuracy of EMVI in rectal cancer. In our study, patients in the EMVI-positive group tended to have statistically higher skewness, kurtosis and entropy of K^trans and maximum of V_e in the EMVI-positive group. Moreover, multivariate logistic regression analysis indicated that K^trans kurtosis and K^trans entropy acted as potential strong predictors of pathological EMVI status. The elevated heterogeneity, originates from more variable cellularity, disordered angiogenesis and variations in necrosis areas, which was proposed as critical characteristic of malignant lesions [36, 37]. Entropy reflects the intratumoral randomness of gray-level distribution. Higher entropy indicates greater heterogeneity, which was reportedly suggestive of higher histological tumor stage and poorer postoperative survival in various malignant tumors, including advanced breast cancer, hepatocellular carcinoma, and rectal cancer [28‐30, 38]. Wilson et al. found the entropy as a potential strong predictor for microvascular invasion in hepatocellular carcinoma [29]. Kurtosis is a measure of the peakedness of the distribution of values in lesions. Wilson et al. also illustrated that greater imaging inhomogeneity might represent histopathological tumor heterogeneity and aggressiveness in hepatocellular carcinoma [29]. Liu et al. proposed that intratumoral heterogeneity might correlate with higher aggressiveness and pathological stages in rectal cancer [30]. Zhu et al. proposed that higher tumor angiogenesis permeability and blood perfusion could facilitate hematogenous metastasis and the formation of EMVI in gastric cancer [39]. Therefore, our results could be explained by the speculation that in the pathological EMVI-positive group, the primary lesions of rectal cancer may be more heterogeneous and aggressive, and local tumor area with hyper-perfusion and high angiogenesis permeability might be more conducive for intravascular tumor cells to protrude outside the intestinal wall, leading to the occurrence of EMVI. These findings realized the assessment of heterogeneity of microcirculation perfusion through signal gray-level distribution characteristics in rectal cancer lesions.

In addition, we evaluated the interobserver variability for HA features extraction. Lambin et al. introduced the radiomics quality score (RQS) in 2017 to ensure the standardization and homogenization of radiomics studies [40]. The aim of the RQS is to evaluate the methodological quality of radiomics-based investigations and identify high-quality results using the ICCs or Cohen’s kappa [41]. Our results indicated good or excellent agreements between the two radiologists for the delineation of whole-VOIs and the calculation of HA features based on multiple slices from DCE-MRI, to ensure the reliability and reproducibility of measurements.

Several limitations of this study should be noted. First, the retrospective nature of this study was a major limitation. Second, the selection of patients with mrEMVI scores of 2 and 3 for further assessment of the diagnostic efficacy of the combined model may have unavoidable selection bias. Third, these patients who received preoperative neoadjuvant CRT were excluded to avoid false-negative in this study. The reliable method of detecting EMVI in patients who undergo neoadjuvant therapy needs further research to confirm. Fourth, the relatively small sample size could compromise the generalizability and stability of our findings. A larger standard multicenter study is needed.

In conclusion, our study demonstrated that HA features obtained from DCE-MRI could help to identify the tumor biology of EMVI and achieve satisfactory radiologic-pathologic matching in rectal cancer patients, even in subpopulations with indeterminate mrEMVI scores of 2 and 3. These findings might be beneficial to preoperative risk stratification and therapeutic decision-making.