In this study, we investigated the diagnostic value of whole-volume ADC histogram analysis in predicting regional lymph node metastasis of periampullary carcinomas. Our results showed that most ADC histogram parameters were significantly different between LNM-positive group and LNM-negative group. At ROC analysis, the tumor size and minimum ADC value generated highest AUC of 0.764. When diagnostic predictive values were calculated with the combined model incorporating tumor size, MR-reported LN status and 75th percentile, the best diagnosis performance was obtained, with AUC of 0.879, sensitivity of 100%, and specificity of 75%.
Due to the region’s anatomical complexity, PCs remain a diagnostic and therapeutic challenge. Preoperative prediction of LNM has received extensive attention due to its prognostic significance and the essential role in patient management. In daily clinical practice, image-based differentiation of metastatic LNs from nonmetastatic LNs mainly depends on morphological features and LN’s sizes, which is inevitably subjective. Moreover, benign LNs with nonspecific inflammatory hyperplasia and metastatic LNs with small sizes also exist. Tseng et al revealed that diagnostic accuracy of conventional CT imaging in assessing PCs’ LNM was less than 81% [24]. In our present study, half of the patients (5 in 10 patients) with LNM were underestimated based on the conventional MR images. Therefore, conventional imaging modalities’ performance in assessment of LN status is inadequate. It is necessary to explore some new method to reveal the histopathological characteristics of metastatic LNs.
Prithivi Raj et al found that PET-CT has a potential ability for detecting PC’ s LNM. In their study, FDG-PET/CT showed a sensitivity of 71.4% and specificity of 77.8% for assessment of LNM with cutoff value SUV max ≥ 2.0. However, small lesions that are less than twice the imaging resolution would yield false negative results because of partial volume effect, and high cost of PET scans may limit its utility in clinical practice [21]. In one previous studies, Bi et al established and validated a radiomics nomogram for preoperatively predicting LNM in PCs. The results showed that the radiomics nomogram incorporating radiomics signature and CT-reported LN status obtained favorable discrimination and calibration ability in both training and validation set, with AUC of 0.853 for each set [22]. However, although the radiomics nomogram performed well, it was complicated and time-consuming, to some extent.
Histogram analysis is a rapidly-emerging noninvasive method in the field of medical imaging, which is capable of quantitatively and objectively assessing tumor heterogeneity, regularity, and image roughness by evaluating the distribution of voxel gray levels without requiring additional invasive procedures [25]. ADC histogram can reflect tissue components with different diffusion features inside the tumor, which was proved to have favorable repeatability and comprehensiveness [20]. LNM has been reported to be associated with ADC histogram parameters of the primary tumor [15–20], which is consistent with our results. ADC values based on water molecular diffusion can reflect the microenvironment of tissues. In general terms, acellular regions (such as cystic and necrotic components) that allow free water diffusion would yield higher ADC values, and highly cellular areas in which water diffusion is restricted by the intensive cell membranes always yield lower ADC values [14, 26, 27]. When the primary tumor progresses, the proliferative ability of the tumor cells and their metastatic capacity were enhanced by the changes of tumor microenvironment [20, 28]. Therefore, patients with LNM showed lower ADC values than that without LNM.
In present study, we not only evaluated various ADC values, but also ADC histogram kurtosis, energy, variance and skewness, which is a more reproducible approach and reflected the distribution of ADC values [29]. In our study, tumors with LNM showed higher kurtosis than tumors without LNM. Kurtosis reflects the peakedness of the histogram distribution and measures the shape of the probability distribution [30]. Higher kurtosis indicates that the distribution tended to have heavier tails or outliers, and the mass of the distribution was more concentrated in tumors with LNM. Previous research has showed that kurtosis of ADC histogram could predict LN status in thyroid cancer [31]. In addition, energy, variance and skewness have also been reported to be correlated with pathological characteristics. Higher energy reflects narrower distribution of intensity levels, higher variance reflects more heterogeneous distribution of ADC values, and higher skewness reflects more asymmetric distribution of ADC values [29, 32]. However, in our study, energy, variance and skewness were not significantly different between LNM-positive and LNM-negative group. Further studies focusing on this topic should be proposed.
For manual tumor segmentation, how to ensure the reproducibility of histogram analysis is the most important issue. Reports suggested that compared with a representative single cross-sectional lesion segmentation, whole-tumor analysis was more reproducible because ADC values were depended on the selected regions of interest, subjective selection of single section would reduce inter- and intraobserver agreements [14]. What’s more, since PCs are often heterogeneous, a single region of interest could not comprehensively reflect different pathological characteristics within the entire tumor. Therefore, in our study, we explored whole-volume ADC histogram analysis, and an overall good interobserver agreement was obtained, with ICCs ranging from 0.766 to 0.967. This was in accordance with previous studies which revealed that whole-volume histogram analyses exhibited favorable repeatability in clinical practice [14, 33, 34].
There were several limitations to our study. First, due to the retrospective design of our study, selection bias inevitably existed even with strict inclusion criteria. Patients with tumors of advanced stage who have lost the opportunity for surgery were excluded. Second, this was a single-center research and the patients number was relatively small, especially the number of patients with LNM, so the diagnostic criteria proposed in our study should be validated in other patient cohort. Finally, we only applied two b values of 0 and 800 s/mm2 for DWI sequence in our study, higher and multiple b values may provide more parameters, which could be further evaluated.