Patients

Patients with RC who underwent surgery between January 2015 and December 2017 at our institution were screened in this retrospective study through reviewing their imaging, clinical, and pathological data and information regarding postoperative treatment and follow-up evaluations. The ethics committee of our institution approved this study and waived informed consent.

The inclusion criteria included: (1) patients with rectal adenocarcinoma who underwent RC surgery at our institution and did not received antitumor treatment before surgery; (2) abdominal PVP CE-CT performed at our institution within 1 month before surgery; (3) initial stage was M0 with diagnosis of RC, and first metastasis appearing in the liver during follow-up; (4) no history of malignancy of other organs; (5) no history of liver disease treatment (except for hepatitis); (6) availability of complete follow-up imaging data obtained during regular follow-up at our hospital; and (7) no contraindications for CE-CT examinations.

The exclusion criteria included: (1) rare pathological types of RC (except for rectal adenocarcinoma), (2) distant metastases or undetermined lesions found before the operation, (3) metastases to other organs before LM, (4) lack of preoperative abdominal PVP CE-CT images, (5) obvious motion artifacts or metal artifacts in CT images, (6) multiple or diffuse benign lesions (such as cysts and hemangiomas) in the liver affecting image segmentation of liver parenchyma; and (7) incomplete clinical indicator or follow-up data.

Two groups were analyzed in this study: (1) the MLM group, defined as RC patients whose metastasis stage was M0 at initial diagnosis, but LM appeared within 24 months after surgery (no other organ metastases occurred before LM), and (2) the non-MLM group, defined as RC patients who showed no metastatic diseases preoperative and within 24 months postoperative images.

Acquisition and scanning parameters for CT imaging

The Toshiba Aquilion, GE Optima CT660, GE Discovery 750 HD, and GE Lightspeed VCT 64-slice spiral CT scanners were used for examinations. The patients fasted for 4–6 hours before the examinations. The examinations were conducted with the patient in the supine position (examinations involving simultaneous abdominal-pelvic CT scans were conducted with the patient in the prone position and with intestinal preparation), both arms raised. The scanning range included at least the entire abdomen. A non-ionic iodine contrast agent (0.3 mg I/mL, Ultravis, Bayer) was injected into the superficial middle elbow vein by using a high-pressure syringe. The dose of the contrast agent was approximately 100 mL with an injection rate of 3.0 mL/s. All patients underwent abdominal scanning with one breath-hold. PVP CE-CT images were acquired with a delay of approximately 65 s after injection of the contrast agent. The scan protocol was as follows: automatic tube current; tube voltage, 120 kV; rotation speed, 0.5 s/rot; screw pitch, 0.984; and slice thickness and layer spacing of conventional scans, 5 mm.

Collection of clinical information

Clinical information of enrolled RC patients regarding sex, age, pathological primary tumor (pT) stage, pathological regional lymph node (pN) stage, hepatitis infection, carbohydrate antigen 19–9 (CA19–9), and carcinoembryonic antigen (CEA) levels were collected through retrospectively reviewing medical records. We also recorded whether postoperative adjuvant treatment, including chemotherapy and chemoradiotherapy.

Therapeutic methods and clinical follow-up

Three antitumor protocols in this study: (1) total mesorectal excision (TME) only, (2) TME followed by adjuvant chemotherapy, and (3) TME followed by adjuvant chemoradiotherapy. None of the patients received antitumor treatment before TME.

All patients were followed up once in every 3 months during the first year, every 6 months in second year, and annually thereafter. Patients without metastases were followed up for at least 2 years after RC surgery. Clinical follow-up after surgery was carried out by reviewing medical records, including serology, colonoscopy, and imaging examinations. Follow-up assessments of LM were mainly based on CE-CT scans. Preoperative images of all patients showed no LM. Among cases showing new suspicious liver lesions during follow-up assessments, some lesions were diagnosed by CE-CT scanning, some uncertain lesions required further diagnoses by liver MRI or positron emission tomography/CT, some were confirmed during follow-up, and some were diagnosed by puncture or surgical pathology. The follow-up time was defined as from the first day after RC surgery to the occurrence of LM or the endpoint of follow-up. The final follow-up was conducted on September 30, 2020.

Image segmentation and radiomics feature extraction

Volume of interest (VOI) segmentation of the whole-liver was performed on PVP CE-CT images by using the open-source imaging platform ITK-SNAP version 3.8 (www.itksnap.org). Image segmentation was manually performed on all and randomly selected cases by two experienced radiologists with 8 and 6 years of experience in abdominal radiology, respectively. They were both blinded to the clinicopathological information. The liver window was adjusted appropriately to optimize the liver parenchyma display (window width, 200–300 HU; window level, 30–70 HU). The VOIs included the whole-liver parenchyma without lesions on the CT images and were manually delineated layer-by-layer, avoiding the edge of the liver (in order to avoid partial volume effects), visible benign lesions in the liver (including cysts, hemangiomas, and calcifications), the main veins and branches in the liver, and hepatic caudate lobe (due to the unclear boundary between the caudate lobe and inferior vena cava in some patients). A schematic diagram of manual segmentation of whole-liver VOI was shown in Fig. 1.

Fig. 1

Schematic diagram of manual segmentation of whole-liver VOI. This was a 70-year-old male patient with RC in the MLM group who developed LM on follow-up images in the 13th month after RC surgery. The red outline in the figure shows the scanning-level area of the liver parenchyma without lesions. Whole-liver VOI without lesions was obtained by sketching the liver layer-by-layer, avoiding the edge of the liver, portal vein, inferior vena cava, and hepatic caudate lobe. A Original PVP CE-CT image of the liver; B manual sketching of one layer; C sketching of one layer was completed; and D schematic diagram after image segmentation

The segmented images were then imported into Artificial Intelligence kit software (A.K. software; version 3.2.5; GE Healthcare, China). A total of 1316 radiomics features were automatically extracted from the PVP CE-CT images, including 18 first-order histogram features, 14 shape features, 24 Gray-level co-occurrence matrix features, 16 Gray-level size-zone matrix features, 16 Gray-level run-length matrix features, 14 Gray-level dependence matrix features, 5 neighboring gray-tone difference matrix features, 186 Laplacian of Gaussian (LoGsigma = 2.0/3.0) features, 744 wavelet features, and 279 local binary pattern features.

Radiomics signature construction and validation

All enrolled patients were randomly divided into training and validation sets in a 7:3 ratio. The training set contained 78 patients (36 patients with MLM and 42 patients without MLM), while the validation set contained 34 patients (16 patients with MLM and 18 patients without MLM).

Interclass correlation coefficient (ICC) values of 30 randomly selected samples were used to compare the consistency of manual segmentation between two radiologists. Features with ICC > 0.75 were selected for subsequent analysis to ensure the high value of the radiomics model. The dimensionality reduction process consisted of two steps. Firstly, the max-relevance and min-redundancy (mRMR) method was used to remove redundant features and retain 20 features most related to MLM. Subsequently, the least absolute shrinkage and selection operator (LASSO) with 5-fold cross validation was carried out to obtain the best feature sets for constructing the radiomics signature. The radscore was calculated by summing the selected features and corresponding coefficients. We then compared the radscores between the two groups in the training and validation sets, respectively.

Construction and validation of prediction models

The clinical model was constructed using the clinical indicators obtained from multivariate analysis. Then, the radscore and significant clinical indicators where p < 0.05 in univariate analysis were included in multivariate logistic regression analysis to construct a combined model. A nomogram was created to make the model visible. Receiver operating characteristic (ROC) curves were used to evaluate the effectiveness of three models in predicting MLM, the area under the curve (AUC), specificity, sensitivity, accuracy, negative predictive value, and positive predictive value were calculated. The DeLong test was used to compare the differences in predicting MLM among the three models. The reliability of the nomogram was determined according to its calibration curve, and the goodness-of-fit was evaluated by the Hosmer–Lemeshow test. Finally, decision curve analysis (DCA) was used to calculate the clinical application value of the three models by quantifying the net benefit at different threshold probabilities. All the procedures for building and validating the radiomics models were shown in Fig. 2.

Fig. 2

Flow chart describing the workflow for construction and validation of the radiomics model

Statistical analysis

Statistical analyses were performed using R version 3.5.1 software and SPSS version 22.0. The Kolmogorov-Smirnov and Levene tests were used to verify the normal distribution and homoscedasticity of continuous variables. The variables with normal distribution and homogeneous variance were compared by two independent-sample t-tests. Otherwise, the Mann-Whitney U test was used. The chi-square test and Fisher’s exact test were used to compare the categorical variables between groups. Variables showing statistically significant differences in univariate analysis were included for further multivariate logistic regression analysis. A two-tailed p < 0.05 was considered statistically significant.