Biparametric magnetic resonance imaging-based radiomics features for prediction of lymphovascular invasion in rectal cancer

Colorectal cancer is a common malignant tumor of the digestive tract. The global morbidity and mortality rates associated with colorectal cancer have been increasing in recent years. This cancer is the second most common cause of cancer-related death, and rectal cancer accounts for one-third of the incidence of colorectal cancer [1, 2]. At present, surgical resection combined with neoadjuvant chemoradiotherapy is the main treatment for rectal cancer [2‐4]. Previous studies have shown that invasion of tumor cells into the microcirculation around the rectum increases the risk of spread and metastasis of rectal cancer. Vascular invasion of rectal cancer is an independent risk factor for postoperative tumor recurrence and low tumor-free survival [5, 6]. A recent study showed that lymphovascular invasion(LVI), rather than the traditionally believed depth of tumor invasion, is even a key risk factor for early rectal cancer metastasis [7]. Therefore, preoperative assessment of LVI of rectal cancer, including identification of small lymphatic and vascular invasion, has very important clinical significance for predicting postoperative recurrence, metastasis, and tumor-free survival of rectal cancer.

High-resolution magnetic resonance imaging (MRI) is the main imaging method for rectal cancer, and it plays a very important role in assessment of the tumor in the preoperative and postoperative periods. Biparametric MRI, including T2-weighted imaging (T2WI) and diffusion-weighted imaging (DWI), has demonstrated reliable results in the detection and staging of rectal cancer [8]. High-resolution MRI shows good performance in the demonstration of larger vessels. In the detection of extramural vascular invasion(EMVI),MRI has moderate sensitivity and high specificity. However, MRI cannot definitively depict vessels with a diameter of < 3 mm, including micro-arteriovenous and lymphatic vessels [9, 10]. Therefore, accurate preoperative imaging evaluation of LVI of rectal cancer is highly challenging.

Radiomics relies on algorithm models to extract quantitative imaging features that cannot be captured by visual inspection. The selected radiomics features can be used to establish predictive model related to disease diagnosis and treatment combined with clinical and pathological data. At present, radiomics as an emergent imaging analysis is widely used in the imaging studies of tumor diagnosis, post-treatment, and prognosis evaluation [11‐13]. In the study of rectal cancer, research has shown that radiomics has good predictive value for nodal metastasis and extramural venous invasion as well as evaluation of neoadjuvant radiotherapy and chemotherapy in patients with rectal cancer [14]. However, relatively few studies have focused on preoperative prediction of LVI of rectal cancer especially in patients with the negative statue of MRI-based extramural vascular invasion(mrEMVI).

The present study was performed to explore the preoperative predictive value and role of biparametric MRI-based radiomic features for LVI of rectal cancer in patients with the negative statue of mrEMVI.

In total, 146 cases of rectal cancer were included in this study (102 cases in the training group, 44 cases in the validation group). Of these 146 cases, 38 were positive for lymphovascular invasion. There were no significant differences in the clinical or pathological results between the groups. The details are shown in Table 2.

The ICCs of the intra-observer and inter-observer reproducibility were satisfactory. Finally,11 and 14 radiomics features were retained from the T2WI and DWI sequences, respectively. 14 characteristic parameters were retained from the combined model. The details are shown in Fig. 3. The Radscore equations were included in Supplemental materials 1 and 2.

The receiver operating characteristics curve showed that the radiomics model based on T2WI and DWI had good predictive performance for lymphovascular invasion of rectal cancer in both the training cohort and the validation cohort. The AUCs of the T2WI model were 0.87 and 0.87 and the AUCs of the DWI model were 0.94 and 0.92. DeLong’s test showed that the DWI model had better performance than the T2WI model (P < 0.05) and that the combined model had better performance than the T2WI model, with AUCs of 0.97 and 0.95 in the training cohort and validation cohort, respectively (P < 0.05). The DWI model had predictive performance comparable to that of the combined model (P > 0.05) (Table 3, Fig. 4). There was an example as shown in Fig. 5. The calibration curves of the combined model demonstrated the best agreement with the ideal curve ( Fig. 6). The decision curves analysis showed that the combined model had the highest standardized net benefit across the range of 0.1–1.0 of the high risk threshold( Fig. 7).

High-resolution MRI is currently the main imaging modality for preoperative and postoperative evaluation of patients with rectal cancer, but it cannot be used to obtain a clear diagnosis of small lymphovascular invasion. The present study was performed to explore the preoperative predictive value of biparametric MRI based radiomics features for LVI of rectal cancer, and the results showed that the radiomics models based on T2WI and DWI had good performance. A recent study of rectal cancer showed that a multiparameter radiomics model had good performance in predicting extramural venous invasion of rectal cancer [20]. In the present study, pathologically confirmed cases of LVI were included. Therefore, the evaluation of microcirculation invasion of rectal cancer was more comprehensively in our study. We selected biparametric images, including those obtained from T2WI and DWI sequences, to delineate lesions in this study for several reasons. On the one hand, the biparametric images are easier to obtain clinically than enhanced images with lower economic cost and a shorter inspection time, which is helpful for multicenter research. Without intravenous injection of gadolinium contrast agent, there is no potential risk of residual contrast agent in the body [21, 22]. On the other hand, previous studies have shown that enhanced T1WI has a limited role in staging of rectal cancer and cannot effectively distinguish the signal intensity of the lesions from that of the rectal wall [23, 24]. T2WI and DWI are the most important sequences for the imaging staging of rectal cancer and are effective in demonstrating the lesions [18, 25]. In addition, the whole-tumor three-dimensional method of delineation was used in this study for two reasons. First, because of the limited resolution of MRI, we could not identify the small-diameter arteriovenous and lymphatic vessels on the images directly. It was difficult to accurately delineate. Second, the three-dimensional method can more accurately and fully characterize the voxel characteristics than the two-dimensional method, as shown in previous studies [26, 27]. In recent years, there have been several studies similarly exploring the role of MRI based radiomics for LVI in patients with rectal cancer [28, 29].But differently, we focused on those patients with negative statue of mrEMVI. As the previous evidence showed, MRI has high specificity but moderate sensitivity in the detection of EMVI. There is no need for patients with positive statue of mrEMVI on the basis of routine MRI evaluation to offer extra imaging method for prediction of lymphovasular invasion.

In this study, the results showed that the AUCs of the radiomics models based on T2WI and DWI in the training cohort were 0.87 and 0.87, and those in the validation cohort were 0.94 and 0.92. The AUCs of the combined model were 0.97 and 0.95 in the training cohort and validation cohort, respectively.The results were better than previous study mentioned above.The probable reasons were as follows. First, the patients in our cohort were all evaluated as mrEMVI(-) preoperately. Consequently, the ratio of early stage patients was higher in our study compared with the previous studies. The cystic and necrotic components in early stage rectal cancer are less than advanced stage.Second, we delineated on biparametric MR images rather than multiparametric or multimodal images,which made the process easier and avoid the possibility of making mistakes in delineation on enhamced images.Third, we ruled out the cases of rectal mucinous adenocarcinoma. Rectal mucinous adenocarcinoma is a rare subtype of rectal adenocarcinoma characterized by lakes of stromal mucin containing scant malignant epithelial cells [30, 31]. Thus, it is difficult to identify the solid component on images. The extent of the tumor can be overestimated. As a result, the delineation of the lesions in our study was easier and more accurate.

Moreover, the discrimination of each model was compared. The results showed that the combined model was better than the T2WI model. Interestingly, the AUC of the DWI model was higher than that of the T2WI model and comparable to that of the combined model. As we know, DWI can detect the motion of water molecules in the extracellular and intracellular spaces in vivo, which reflects the cell membrane integrity and cell density [32]. Previous studies have shown that high b-value DWI can improve the accuracy of imaging staging of rectal cancer and has a role in distinguishing tumor foci and fibrous reactions after neoadjuvant chemoradiotherapy [33‐35]. In rectal adenocarcinoma, although the component of necrosis and cysts can increase the diffusion, the solid component consisting of the tumor cells always causes diffusion to become restricted. Consequently, the region of tumor cells demonstrates much higher signal intensity than the rectal wall on DWI, which is helpful in detecting the tumor regions and facilitating more accurate delineation. Additionally, the regions of interest on DWI may better characterize tumors and include more biological information than T2WI because DWI has the capability to reflect the tumor microenvironment. Furthermore, the calibration and clinical applicability of the models were taken into consideration. The results showed the combined model demonstrated the best degree of calibration and clinical applicability. The results showed that the combined model can effectively predict LVI of rectal cancer in patients with the mrEMVI (-). DWI was helpful in discriminating the status of LVI in this study.

This study had some limitations. First, the sample size was not large enough. Large-sample multicenter studies are still needed. Second, patients who underwent preoperative neoadjuvant radiotherapy and chemotherapy were excluded. There will be selection bias in this cohort of patients.Third, because of the limited sample size, cases of vascular and lymphatic invasion were not separately divided into groups. Fourth, this was a retrospective study. Finally, the patients’ clinicopathological data was not included in the predictive model.

In conclusion, the radiomics model based on bipapametric MRI showed good preoperative predictive value for lymphovascular invasion of rectal cancer with the mrEMVI (-). Radiomics from bipapametric MRI has the potential to serve as a noninvasive and effective biomarker to predict the status of lymphovascular invasion in rectal cancer preoperatively.