A nomogram based on dual-layer spectral detector CT-derived 40KeV virtual monoenergetic images for preoperative prediction of simultaneous distant metastasis in colorectal cancer.

Introduction
Colorectal cancer (CRC) is currently the third most common malignant tumor in the world in terms of incidence rate and the second most common in terms of mortality rate, and the incidence is increasing year by year, and Asia is the region with the highest incidence (51.8%) and mortality rate (52.4%) both in males and females [1, 2]. Recurrence and metastasis are leading causes of death in CRC patients [2, 3]. With the gradual popularization of preoperative neoadjuvant radiotherapy in the treatment paradigm of locally progressive CRC, the rate of local recurrence has significantly decreased, but distant metastasis remains the leading cause of death in CRC patients [4–6]. Distant metastases of CRC are categorized into synchronous distant metastasis (SDM, distant metastases found before or at the time of diagnosis of lesions at the primary site) and metachronous distant metastasis (MDM, distant metastases found after radical surgery) [7]. Previous studies have found that SDM occurs in 20–34% of CRC patients at the time of diagnosis, with the liver being the most common site of metastasis, followed by the lungs, peritoneum, and distant lymph nodes [8–11]. The principles of management are different for SDM and MDM, and the management of metastases at different sites is also different. Relevant guidelines have stated that SDM requires consideration of the primary focus, and when occurrence at the primary focus is associated with emergencies such as obstruction and significant bleeding, treatment of lesions at the metastatic focus should be considered after management of cancer at the primary focus [11, 12]. Surgical resection has been recognized as the therapy of first option for focal CRC distant metastases because of the significant improvement in survival, and patients with CRC who have surgical resection of hepatic or pulmonary metastases have a significantly higher 5-year survival rate [13–15]. Therefore, the correct preoperative identification of SDM is crucial for the subsequent development of treatment strategies in CRC patients, and noninvasive imaging assessment methods are indispensable and important tools for accomplishing it.
Imaging methods for CRC distant metastasis mainly include thoracic and abdominal CT, abdominal MRI and PET-CT. MRI has become the first choice for baseline screening of rectal cancer, and MRI enhancement of the liver is commonly used for the screening of liver metastases. The colon is located in the abdominal cavity with high mobility and is easily affected by respiratory movement, MRI comes with limitations in colon imaging, and enhanced CT examination is often used for baseline examination of colon cancer and evaluation of systemic lesions. However, some pulmonary nodules are often found during chest CT examination in the general assessment of CRC patients, and chest CT alone cannot determine whether they are metastatic nodules. In addition, due to the subjective factors of radiologists, there are sometimes some misdiagnoses and missed diagnoses [16, 17]. PET-CT is more expensive and tends to have lower diagnostic specificity. Therefore, patients with CRC need a better baseline examination and systemic evaluation method before surgery to provide reference for subsequent precision treatment.
Dual-layer spectral detector computed tomography (DLSCT) is an energy-spectrum CT based on a dual-layer detector that obtains multiple functional images such as effective atomic number images (Eff-Z), virtual monoenergetic images (VMI), iodine concentration (IC) images and extracellular volume fraction (ECV) iamges, in a single scan while not increasing the radiation dose [18]. It was found that among the various energy levels of VMI, the energy level of 40 keV was more effective in displaying CRC lesions and could significantly improve the diagnostic performance of CRC [19]. Radiomics is the extraction and analysis of a massive number of high-dimensional, quantitative image characteristics from images, combining methods such as quantitative analysis of image images and machine learning to provide more accurate information [20]. To the best of our knowledge, there are few reports on CRC based on DLSCT radiomics analysis [21–24], and there has not been any study on its prediction of SDM in CRC. Therefore, the aim of our study was to develop and validate a nomogram based on clinical features, spectral parameters and radiomics features of enhanced three-phase DLSCT-derived 40KeV-VMI for preoperative prediction of SDM in CRC.

Methods

Patients
This is a retrospective multicenter study, and our study protocol was approved by the ethical review boards of two hospitals without obtaining written informed consent.
This study retrospectively collected a preliminary group of 238 CRC patients seen at Hospital A and Hospital B from June 2022 to April 2023. We collected clinical, pathologic, and DLSCT imaging data from these patients. SDM was confirmed by postoperative histopathological examination, imaging examination and clinical follow-up. The inclusion criteria were as follows: (1) pathologically confirmed colorectal cancer; (2) abdominal and pelvic spectral-enhanced CT scan performed within 2 weeks prior to surgery; (3) no history of other tumor-related conditions. The exclusion criteria were as follows: (1) confirmed by pathology as a nonordinary adenocarcinoma, such as mucinous adenocarcinoma; (2) small lesions or the presence of metal stents in the diseased bowel lumen resulting in lesions that could not be identified or outlined; (3) incomplete clinical and pathologic information or patients who were lost to follow-up; (4) neoadjuvant therapy prior to DLSCT examination. Finally, 137 CRC patients [SDM+ (n = 51); SDM−(n = 86)] who met the above requirements were included in this study. We divided the 137 CRC patients who met the requirements into a training group (n = 90, from Hospital A) and an external validation group (n = 47, from Hospital B). Figure 1 shows the flowchart of enrolled patients.

DLSCT image acquisition
The scanning equipment was a DLSCT instrument (IQon Spectral CT; Philips Healthcare, Best, The Netherlands). All patients were placed in the supine position, head first, with both upper limbs on either side of the head, and scanning started at the apex of the diaphragm and proceeded to the lower edge of the anal canal. The contrast agent iohexol (350 mg/ml) was injected via an anterior elbow vein using a double cylindrical hyperbaric syringe at a dose of 1.5 ml/kg and an injection rate of 3.0 ml/s. The threshold monitoring method was adopted, with the threshold set at 150 HU. When the threshold is reached, an instruction to inhale and hold your breath will be activated first (default duration is approximately 5 s), followed by the acquisition of arterial phase image. The venous phase (VP) image was acquired after 60 s, and the delayed phase (DP) image was acquired after 180 s. The DLSCT scanning parameters included the following: tube voltage, 120 kv; tube current, automatic tube electric flow controlling method; helical pitch, 1.234; ball tube rotation time, 0.75 s/r; and collimator width, 64 × 0.625 mm. The acquired three-phase enhanced images were reconstructed with holographic spectral-based image (SBI) to obtain spectral images. All reconstructed layer thicknesses and layer spacings were 1 mm. All scan parameters of DLSCT were consistent between the two hospitals.

DLSCT quantitative parameter measurements
On the specialized postprocessing workstation for DLSCT (IntelliSpace Portal, 12.1, Philips Healthcare), we used Spectral CT viewer software to analyze three-phase enhanced images by reconstruction. The circular ROI is placed in the most significantly enhanced solid area of the lesion, avoiding intestinal contents, fat, cystic lesions, necrotic areas, blood vessels, and calcifications [24, 25]. The ROI is located on the slice with the largest axial diameter of the tumor, as well as the adjacent upper and lower slices. The average value of the three measurements is calculated for subsequent analysis. The size of the ROI on different phases is maintained consistent using the copy and paste method. The ISP post-processing workstation can directly switch between different energy spectrum sequence images such as IC and VMI for reconstruction, which also ensures that the size and position of the ROI are consistent in different sequence images. Additionally, when delineating the same layer of the tumor in the delayed phase, we place a circular ROI in the aorta or iliac artery, and be careful to avoid atherosclerotic plaques on the vessel wall. The most recent hematocrit (HCT) within 1 week prior to DLSCT examination was recorded for all patients, and the ECV was calculated using the formula ECV (%) = (1-HCT) × (ICtumor/ICaorta) × 100%, where ICtumor and ICaorta are ICs in the DP of the tumors and the aorta or iliac artery at the same level, respectively (Supplementary Material S1).
A radiologist with 5 years of abdominal diagnostic experience, blinded to clinical and pathological information, performed two quantitative measurements with an interval of one month between each measurement. The contoured results were reviewed and revised by a radiologist with 10 years of abdominal diagnostic experience. The average of the two measurements was calculated for subsequent analysis.

Segmentation of DLSCT images
Manual segmentations of CRC regions were performed using ITK-SNAP software (version 3.8.0, http://www.itksnap.org) for AP, VP, and DP images, respectively. We delineated the contours of the CRC lesions layer by layer to obtain the volume of the entire primary lesion, which is also known as the three-dimensional region of interest (VOI). During the delineation process, we adjusted the window width and window level simultaneously to review coronal and sagittal images to help identify the boundaries of the tumor [26]. The segmented VOI avoided obvious visible blood vessels, fat, and air cavities. All images were manually segmented by a radiologist with 5 years of experience in abdominal diagnosis. The segmentation results were then reviewed and revised by another radiologist with 25 years of experience in abdominal diagnosis, who was unaware of the clinical and pathological conditions of the patients (Supplementary Material S2).

Preprocessing of clinical variables
Clinical information included age, gender, tumor location, clinical T-stage(cT-stage), clinical N-stage(cN-stage), carcinoembryonic antigen (CEA) level, Carbohydrate antigen 19 − 9 (CA19-9) level, HCT, Neutrophil-to-lymphocyte ratio (NLR), platelet-to-lymphocyte ratio (PLR), and systemic immunoinflammatory index (SII) within 1 week prior to the DLSCT examination. Two radiologists with 5 and 10 years of experience in abdominal diagnosis, respectively, observed the enhanced DLSCT images of the abdomen and pelvis without knowledge of clinical and pathological information. Clinical T-stage and N-stage of CRC were evaluated with reference to the Colorectal Cancer Staging System of the American Joint Committee on Cancer Staging (AJCC), 8th edition [27]. Disagreements were resolved by a third radiologist with 25 years of experience in abdominal diagnosis.

Extraction and selection of radiomics features
The open-source Python package (PyRadiomics version 3.0) was applied to extract the radiomics features, including shape features, first-order intensity features and texture features, of the VOI regions of 40KeV-VMI in AP, VP and DP, respectively. At the same time, VOIs were filtered to obtain more kinds of radiomics features.
To reduce the model complexity as well as to prevent the overfitting problem, the radiomic characteristics extracted from DLSCT images of enhanced three-phase were subjected to dimensionality reduction. The final selected radiomics features of the three-phase enhanced 40KeV-VMI were standardized and used for subsequent analysis (Supplementary Material S3).

Construction and utility evaluation of the predictive model
Univariate and multivariate binary logistic regression analysis were used to screen the clinical characteristics and spectral quantitative parameters that were significantly correlated with SDM of CRC to construct a clinical model. The AP, VP and DP radiomics models were established, and radiomics scores (Radscores) were calculated by linear combinations of the corresponding radiomics features. A visualized nomogram was built by integrating the radiomics scores, clinical features and spectral quantitative parameters of 40KeV-VMI in the enhanced three-phase. The receiver operating characteristic curve (ROC) was used as an evaluation index for the predictive efficacy of the five models for SDM of CRC, and the sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV) and area under the curve (AUC) were calculated.
The Hosmer-Lemeshow test was utilized to evaluate the goodness of fit of predictive models. The calibration performance and clinical utility of the models were also assessed by plotting calibration curves and decision curve analysis (DCA).

The determination criteria of SDM for CRC [28, 29]
SDM is diagnosed through imaging examinations, confirmed through clinical follow-up or postoperative pathological examination. Imaging examinations: (1) When there are no other confirmed tumors in other parts besides the primary site of CRC, a single or multiple space-occupying lesions outside the primary site are found during the baseline examination, and CT/MRI/PET-CT all consider them as metastatic lesions; (2) Changes in metastatic lesions after treatment without surgery or preoperative neoadjuvant radiotherapy. Clinical follow-up is completed through querying the inpatient electronic medical records and telephone inquiries.

Statistical analysis
Quantitative data that corresponded to a normal distribution are shown as the mean ± standard deviation, and comparisons between two groups were performed through independent samples t tests; quantitative data that did not correspond to a normal distribution are shown as M (Q1, Q3), and comparisons between two groups were made through the Mann-Whitney U test. Count data are shown as frequencies (n) and percentages (%). The DeLong test was used to compare the performance of each model. Heatmaps were plotted using software (HemI v1.0) to demonstrate the screened radiomic characteristics of the enhanced three-phase images. The algorithm modeling process was performed on the InferScholar research platform (version 3.5, Beijing, China). Statistical analysis was applied in SPSS for Windows (version 27.0). Statistical significance was determined at P < 0.05 (two-tailed).

Results

Clinical and spectral quantitative characterization of the study population
The clinical and spectral quantitative characteristics of the 137 CRC patients in the training and external validation groups are demonstrated in Table 1. Fifty-one CRC patients had SDM, including 18 cases of liver metastasis, 6 cases of lung metastasis, 12 cases of distant lymph node metastasis, 4 cases of simultaneous liver and lung metastasis, and 11 cases of multiple distant metastasis. Fifteen cases were confirmed by surgical pathology, and the rest were confirmed by imaging examination and clinical follow-up.
The difference in the incidence of SDM between the training and external validation groups was not statistically significant (35.6% in the training group and 40.4% in the external validation group, P = 0.58). Differences in tumor location, ICAP, CT40KeVAP, ICVP, and PLR were statistically significant between the two cohorts (all P < 0.05).

Selection of clinical variables
CEA (P = 0.012) and ECV (P = 0.002) were independent risk factors for predicting SDM in CRC (Table 2).

Extraction and selection of radiomic features
1746 radiomics features were extracted from 40KeV-VMI of 137 CRC patients in AP, VP and DP, respectively (Supplementary Material S3). Five AP radiomics features, six VP radiomics features, and five DP radiomics features were finally screened by LASSO dimensionality reduction process (Fig. 2). Heat maps of key radiomics features were plotted based on standardized radiomics feature values (Supplementary Material S4).

Development of the predictive model
CEA and ECV were constructed as clinical models. Based on the radiomics features screened by LASSO dimensionality reduction processing, the enhanced three-phase radiomics models were constructed separately and the Radscores were calculated for each CRC patient (Supplementary Material S5). A nomogram was constructed by integrating the three-phase Radscores and the clinical risk factors (Fig. 3).

Comparison of models and evaluation of clinical utility
The predictive performance of the five models for SDM of CRC in the training and external validation cohorts was evaluated by ROCs (Fig. 4). In the training group, the AUC of the nomogram was 0.938 (95% CI: 0.866–0.978) and was significantly better than that of the three-phase radiomics models (all P < 0.05). In the external validation group, the AUC of the nomogram was 0.930 (95% CI: 0.817–0.984) and was remarkably superior to that of the DP model (P = 0.040) and the clinical model (P = 0.044) (Table 3).
The calibration curves of all models in the training group and external validation group showed good agreement between the five models predicting the probability of SDM in CRC and the actual observed rate (Supplementary Material S6). The Hosmer-Lemeshow test results showed that there was no significant statistical difference between each model and the nomogram in the training and external validation groups, which indicated that there was no remarkable deviation from the ideal fit (Supplementary Material S7). The nomogram had a better critical net benefit than the other four models in predicting SDM of CRC within most threshold probabilities (Fig. 5).

Discussion
In our study, a nomogram integrating radiomics features of enhanced three-phase, clinical features, and spectral quantitative parameters of 40KeV-VMI was constructed for preoperative prediction of SDM in CRC. The results showed that the nomogram had the highest AUC compared to a single clinical or radiomics model and its net clinical benefit for predicting SDM was higher than that of the other four models. This suggests that the nomogram based on DLSCT-derived 40KeV-VMI of enhanced three-phase could be used as a preoperative adjunct to improve the diagnostic accuracy in patients with suspected SDM, helping to improve clinical decision-making and the development of individualized treatment strategies. CEA and ECV were found to be independent risk factors for SDM in CRC by binary logistic multifactorial regression analysis. Several previous studies have shown that serologic markers such as CEA and CA19-9 can be used to predict distant metastasis in CRC patients [30–32]. NLR and SII are important inflammatory indicators related to tumor proliferation and metabolism, which are closely associated to metastasis and poor prognosis of CRC [33–35]. Our results showed the same finding that patients with high CEA, CAl9-9, PLR and SII had a greater chance of developing SDM, although only CEA was significantly different after multifactorial regression screening, which may be related to the CT scanning protocols and patient characteristics and needs to be further investigated by increasing the sample size in the future. ECV is the sum of intravascular space and extracellular extravascular volume fraction; it can reflect both the microvascular density (MVD) and the degree of stromal fibrosis of the tumor, and thus can more comprehensively reflect the microenvironment of tumor growth. In the microenvironment of tumor growth, tumor neovascularization and extracellular matrix are important factors in the survival and growth of malignant tumors and playing a crucial role in their invasion and metastasis. CRC is a rich blood supply tumor, and there are a large number of new but immature blood vessels in the tumor. These new blood vessels have thin walls and large blood vessel fragility, which makes the blood flow and vascular permeability are large, and it is easy for tumor cells to fall off into the circulation, thereby increasing the possibility of distant metastasis [36]. Previous studies have shown that MVD may be used as a marker for assessing tumor angiogenesis [37], and stromal fibrosis and high MVD in CRC are significantly associated with poor prognosis [25, 38, 39]. In our study, ECV was significantly different between the SDM and non-SDM groups, and ECV was an independent risk factor for SDM of CRC, which is also in line with previous studies.
Each tissue has its own unique distribution of attenuation values. Conventional CT is based on the Hounsfield unit to distinguish different tissues, but the overlap of attenuation between many tissues often limits this distinction. In contrast, the simultaneous acquisition of energy by the low-energy and high-energy dual-layer detectors of DLSCT can be used to identify different tissues by their energy-dependent attenuation characteristics and to obtain relevant information such as tissue density and atomic number, which can be used to distinguish different tissues that cannot be distinguished by conventional CT under similar attenuation coefficients [40]. Previous studies have found that compared with conventional CT, DLSCT-derived 40KeV-VMI has higher signal-to-noise ratio and contrast-to-noise ratio, it can more clearly show the enhancement of colorectal wall, which can significantly improve the diagnostic performance of CRC [19]. Nagayama et al. found that 40KeV-VMI had superior detection and significance for liver metastases with a diameter of less than 1 cm [41].
However, a single spectral quantitative parameter value measured by ROI may not fully reflect the whole tumor heterogeneity [23]. Radiomics integrates many of the high-dimensional imaging features used to quantify tumor heterogeneity and has been widely used in various studies of tumors; in addition, 3D VOIs are more representative of the heterogeneity of the entire lesion than 2D VOIs [42, 43]. Previous studies have found that DLSCT-based radiomics features have good diagnostic efficacy in preoperative prediction of lymph node metastasis and tumor deposition in CRC [23, 24].
In our study, radiomics analysis based on DLSCT-derived 40KeV-VMI was applied to preoperative prediction of SDM in CRC. The radiomics features used for modeling in our study primarily include the Gray-Level Co-occurrence Matrix (GLCM), Neighborhood Gray-Level Difference Matrix (NGTDM), Gray-Level Dependency Matrix (GLDM), Gray-Level Run Length Matrix (GLRLM), and Gray-Level Size Matrix (GLSZM). Among these, GLCM_Correlation reflects the linear dependency between the gray values of adjacent pixels, with lower correlation indicating tumor texture heterogeneity. NGTDM_Contrast, NGTDM_Busyness, GLSZM_ZoneEntropy, and GLCM_Maximum correlation coefficient are all important parameters reflecting texture complexity. As these values increase, texture complexity increases, indicating heightened heterogeneity in tumor internal structure, uneven distribution of tumor density, disrupted tissue structure, and irregular vascularization. These factors are positively correlated with tumor invasiveness, metastasis, and recurrence [44–46]. GLDM_SmallDependenceLowGrayLevelEmphasis and GLSZM_LargeAreaLowGrayLevelEmphasis reflect the distribution frequency of low gray level regions; GLRLM_LowGrayLevelRunEmphasis represents the continuity of low gray level regions in the image. The presence of necrotic or low cell density regions within the tumor indicates poor differentiation, severe hypoxia, and a strong tendency toward invasion and metastasis [47, 48]. These texture features can reflect the spatial relationships between pixels in the image, thereby providing a more comprehensive and in-depth description of tumor heterogeneity [49, 50]. The nomogram integrating clinical characteristics, spectral quantitative parameters, and enhanced three-phase 40KeV-VMI radiomics features demonstrated the best predictive performance in SDM, outperforming single clinical models and radiomics models. This can be understood as the nomogram model incorporates the information provided by the 40KeV-VMI, comprehensively capturing the heterogeneity within the tumor. This nomogram model can serve as a reliable auxiliary tool for non-invasive preoperative prediction of SDM.
Despite the good predictive performance of our nomogram, this study is affected by some limitations that should be acknowledged. First, although two central studies were conducted, the sample size of the SDM positive group was limited, and the total sample size was also relatively small. This might lead to overfitting of specific sample characteristics during the model training process. Moreover, the insufficient representativeness of the samples may cause bias in the model during the training stage, resulting in insufficient generalization. In the future, a larger sample size is needed, and combined with multi-center and cross-population data, the model should be verified to ensure its generalization ability. Secondly, distant lymphatic metastasis was not discussed separately from other distant metastases in this study, and the sample size can be expanded to further discuss its differences in the future. Thirdly, the spectral quantitative parameters included in this study are limited. In the future, while increasing the sample size, we will further incorporate more spectral parameters from the delay period for analysis and research. Fourthly, this study did not analyze the conventional mixed energy images, we can compare and analyze the 40KeV-VMI reconstruction method derived from DLSCT in the future. Finally, creating the manually outlined 3D VOIs were time-consuming and laborious with unavoidable bias, and a tool based on artificial intelligence automated segmentation is needed to improve segmentation efficiency and accuracy.

Conclusion
In conclusion, the nomogram incorporating radiomics features of DLSCT-derived 40KeV-VMI, spectral parameters and clinical features showed excellent predictive performance in preoperatively predicting SDM in CRC, which can help clinicians make accurate individualized treatment plans.

Supplementary Information
Below is the link to the electronic supplementary material.