Background
Cholangiocarcinoma (CCA) is a fatal hepatobiliary adenocarcinoma with a high mortality rate and poor prognosis, accounts for almost 3% of all gastrointestinal malignancies [
1]. Extrahepatic cholangiocarcinoma (ECC) is highly invasive, difficult to diagnose early, and hard to surgically remove due to its complex anatomy. As a result, patients has 5-year relative survival rates ranging from 2 to 30% [
2]. While surgical resection is the primary treatment option for ECC patients, the recurrence rate remains disappointingly high [
3,
4]. In this context, vigilant postoperative monitoring is crucial for assessing treatment efficacy and the potential for recurrence [
5]. However, patients who have not received appropriate treatment may be subjected to excessive or inadequate treatment, causing unnecessary side effects [
6]. Given these challenges, he preoperative identification of prognostic biomarkers is imperative to assist patients in selecting the most appropriate treatment strategy prior to the commencement of therapy, thereby enhancing the likelihood of a favorable prognosis. Based on this, constructing a predictive model using preoperative biomarkers can provide a basis for clinicians to develop personalized treatment strategies.
Tumor-node-metastasis (TNM) staging serves as a valuable reference for clinicians in prognostic assessments and treatment planning [
7]. However, the TNM staging system solely considers tumor size, depth, and lymph node involvement, disregarding the overall patient condition and other prognostic factors. Consequently, it fails to comprehensively depict the biological behavior and prognosis of tumors, resulting in disparate prognoses for certain tumors within the same stage while others may exhibit similar prognoses across different stages [
8]. Therefore, relying solely on three criteria—tumor size, lymph node metastasis, and distant metastasis—to judge prognosis is imprecise and inaccurate. As a non-invasive diagnostic tool, MRI has been widely used in the diagnosis and treatment of ECC [
9]. MRI can provide high-resolution images that can display the location, shape, size, and infiltration range of ECC, which plays an important role in preoperative evaluation, selection of surgical plans, and evaluation of treatment effects and recurrence [
10]. Additionally, MRI can also be utilized for assessing the correlation between lesions of bile duct cancer and adjacent tissues or vital structures. By comparing the discrepancies in MRI signals between normal surrounding tissues of the bile duct and those surrounding the lesion of cholangiocarcinoma, it becomes feasible to determine whether there is infiltration into the neighboring tissue by the lesion, thereby aiding in evaluating the extent of tumor invasion and metastasis [
10,
11]. In recent years, the utility of MRI as a prognostic tool for postoperative survival in ECC patients has increasingly come into focus [
12,
13].
MRI is a crucial modality for assessing signal alterations in the liver and spleen. Research has demonstrated a close association between abnormal liver MRI signals and liver injury, iron overload, and anemia [
14]. Specifically, the intensity of splenic signals on MRI has been shown to have a substantial association with serum ferritin levels, reflecting the severity of anemia. Iron overload can result in impaired organ function, potentially leading to organ failure [
15]. Moreover, certain biomarkers indicative of liver injury and cholestasis, such as gamma-glutamyl transferase (GGT), alanine aminotransferase (ALT), aspartate aminotransferase (AST), and total bilirubin (TBIL) have been utilized to reflect the status of hepatic damage and evaluate prognostic outcomes in patients with cholangiocarcinoma.
In general, this study endeavors to formulate a preoperative prognostic model, integrating preoperative laboratory analyses and MRI characteristics, to precisely forecast the prognosis for ECC patients undergoing curative resection.
Materials and methods
Patients
This study was conducted in strict adherence to the principles set forth in the “Declaration of Helsinki.” The Institutional Review Board in The Affiliated Hospital of Southwest Medical University granted approval for this retrospective study, and the requirement for written informed consent from the patients was waived in accordance with the regulations. To ensure the confidentiality of patient information, all identifying details were meticulously removed from the records. Patients with ECC diagnosed by postoperative pathology at our hospital from January 2013 to July 2021 were included. The inclusion criteria mandated that patients (1) had undergone radical resection or pancreaticoduodenectomy, (2) were free of other malignant neoplasms or intrahepatic metastases, (3) had no prior exposure to anticancer treatments like radiotherapy or chemotherapy before surgery, and (4) had undergone an MRI scan within one month prior to surgery. Conversely, exclusion criteria included (1) death during the perioperative period; (2) incomplete clinical and imaging data; and (3) lesions smaller than 5 mm and undetectable in MRI. Ultimately, a total of 104 patients satisfying the stringent criteria were enrolled in this study (Supplementary Fig.
1). To ensure accurate and standardized pathological determinations, an experienced abdominal pathologist with a decade of diagnostic expertise meticulously reviewed the pathological assessments of ECC. In addition to the pathological evaluation, various clinical parameters were taken into consideration, encompassing gender, age, ALT, AST, TBIL, direct bilirubin (DBIL), GGT and Carbohydrate antigen199 (CA199).
MRI scan
The MRI images of all patients included in this study were obtained using a 3.0T MRI scanner (Acthia3.0T, Philips) and a 16-channel abdominal coil. All patients were instructed to fasting for 6-8 h before examination, and were requested to practice breathing and breath holding. The scan range was from the upper edge of the liver to the descending segment of the duodenum. The following MRI sequences were collected: axial fat-suppression turbo spin echo (TSE) T2-weighted imaging (T2WI) sequence and axial diffusion-weighted imaging (DWI). This study mainly aimed at analyzing the imaging features of T1WI, T2WI, DWI sequences, and the apparent diffusion coefficient (ADC) maps. The parameters of some MRI sequences are presented in Supplementary Table
1.
The ADC maps were automatically derived from DWI (b = 0 and 800 s/mm2) images in Extended MR Workspace R2.6.3.1 (Philip Healthcare). The signal intensity of the lesion on the DWI (b = 800 s/mm2) sequence was measured, and the ADC values for the lesion (SIlesion) on the ADC map were similarly determined. The ADC values were measured by copying the same regions of interest (ROI) from the DWI sequence to the ADC map. The ROI of the lesion should be near the center of the lesion and measured thrice to obtain the average. During measurement, care should be taken to avoid artifacts, blood vessels, bile ducts, and necrotic areas. The ADC values for the spleen (SISpleen) and liver (SILiver) were measured in the same way as those for the lesion. In addition, the signal intensity ratio of the liver-to-muscle (SIRLiver/Muscle) was measured using images obtained with each T2-weighted image by means of the ROI positioned in the liver and paraspinal muscle. In the same way, SIRSpleen/Muscle was also determined on the T2WI sequence.
MRI features
All the scan sequences of the patients were comprehensively observed, and the following imaging traits were collected: (1) lesion’s location (Perihilar cholangiocarcinoma (pCCA) is localized between the left/right hepatic duct and insertion of the cystic duct into the common bile duct and distal (dCCA) is confined to the common bile duct); (2) morphology (mass-forming type, periductal infiltrating type, and intraductal growth type); (3) tumor size (maximum diameter of the tumor measured on the axial T2WI images); (4) lesion’s signal (homogeneous or heterogeneous); (5) intrahepatic bile duct dilatation (diameter of the widest intrahepatic bile duct at 1 cm from the confluence of the left and right hepatic ducts was measured); (6) DWI signal (when b = 800 s/mm2, the lesion showed high signal or low signal); (7) SILesion; (8) SISpleen; (9) SILiver; (10) SIRLiver/Muscle; 11) SIRSpleen/Muscle.
Follow-up
Patient outcomes, including survival time and cause of death, were obtained through medical records, telephone follow-up, or local population database. Overall survival (OS) was defined as the duration between the date of operation and the date of death or the last follow-up. It is noteworthy that the last follow-up for this study occurred on July 31, 2021. Within the cohort, patients were categorized into two distinct groups: the death group, comprising individuals who passed away due to ECC during the follow-up, with survival duration measured from surgery onset to death date; and the survival group, consisting of those alive at follow-up’s end, with survival time calculated from surgery commencement to the last outpatient visit. It is important to note that the survival group data was considered as censored data, allowing for a comprehensive analysis of patient outcomes, even beyond the study’s timeframe.
Statistical analysis
Continuous variables adhering to a normal distribution were articulated as mean ± standard deviation and compared using the t-test. Conversely, variables deviating from normal distribution were represented as median and interquartile range (25th and 75th percentiles, Q1, Q3) and analyzed via the Mann-Whitney U test. Categorical variables were expressed as counts and percentages, with the chi-square test employed for comparative analysis. Statistical significance was established at a two-tailed p-value of < 0.05.
Univariate COX regression analysis was performed to identify potential risk factors for OS. Subsequent to univariate analysis, variables with
p < 0.10 were entered into multivariate analysis. The independent variables were tested for proportional hazards assumptions and applied to model development. The Akaike information criterion was employed to choose the final predictive model. COX regression analyses were used to detect the independent prognostic factors for OS. A two-tailed
p-value < 0.05 were considered as statistically significant in univariate or multivariate analysis. Additionally,, the hazard ratio (HR) and the associated 95% confidence interval (CI) were also calculated. In R Studio, independent risk factors were used to establish a nomogram for predicting 1- and 2-year OS. Then, the internal verification of the nomogram was performed using 2000 bootstrap iterations. Calibration curves and C-index were used to evaluate the precision of the nomogram, and ROC curves were used to evaluate the accuracy of the nomogram’s 1- and 2-year forecasts. Besides, k-fold cross-validation was used to verify the stability of the nomogram in predicting OS. All statistical analyses were implemented using SPSS (version 26.0) and R software (version 4.1.2,
https://www.rproject.org). To evaluate and validate the reliability of the measurement data, a subset of 20 patients was selected from a total cohort of 104 patients. Two experienced radiologists were invited to assess the MRI features across T1WI, T2WI, and DWI in order to compare inter-observer consistency. By calculating the intraclass correlation coefficient (ICC) or Kappa values for each feature among observers, we confirmed the reliability of acquiring MRI feature data. Notably, an ICC or Kappa values ≥ 0.7 indicated excellent agreement among observers.
Discussion
ECC, a relatively uncommon biliary tract malignancy, is characterized by its insidious onset, complicating early diagnosis and often leading to advanced-stage presentations. This typically impedes surgical resection in most cases, contributing to a dismal prognosis [
16], thereby highlighting the imperative for early detection. Consequently, precise prognostication of ECC patient outcomes is vital for guiding treatment strategies and enhancing quality of life. Our research introduces a novel nomogram that combines preoperative clinical variables and traditional imaging features to predict prognosis. Distinct from radiomics approaches, this nomogram allows clinicians to estimate predictive probabilities without the need for specialized software or region-of-interest (ROI) delineation, simply by evaluating traditional imaging features. This enables quick assessments based on each patient’s specific situation. Our study identified key prognostic indicators, including gender, ALT, GGT, DBIL among clinical features and tumor size, tumor location, SIR
Liver/Muscle ratio, SIR
Spleen/Muscle ratio among MRI features. Importantly, the incorporation of SIR
Liver/Muscle and SIR
Spleen/Muscle ratios as predictors represents a novel discovery not previously documented in cholangiocarcinoma prognosis studies.
Specifically, we observed a negative correlation between SIR
Liver/Muscle and the risk of death. When the muscle signal remains unchanged, the decrease in SIR
liver/muscle is caused by a decrease in liver signal. Previous studies have linked diminished T2-weighted imaging (T2WI) signal on MRI to abnormal iron deposition in the liver, a primary storage site for iron [
17]. Liver cell damage leads to the leakage of stored ferritin into the bloodstream [
18], markedly lowering T2WI signal intensity [
19]. Additionally, malignant tumors can produce and release ferritin, leading to anemia and blood overload [
20]. Research has demonstrated that tumors associated with pernicious anemia portend a poorer prognosis [
21]. Thus, the decrease in SIR
liver/muscle indirectly reflects abnormal liver iron metabolism and pernicious anemia, implying a poor prognosis. Similarly, the spleen plays a crucial role in blood circulation, with iron transported by erythrocytes accumulating in the reticuloendothelial cells of both the liver and spleen, influencing iron metabolism [
22]. In the context of anemia or blood overload due to malignancy, splenic function is also impacted. A strong correlation between ferritin levels and abnormal spleen MRI signals has been established [
17]. MRI evaluation of iron overload is feasible, as it reveals a reduced spleen signal intensity relative to paraspinal muscle in T2WI sequences [
23]. In our study, patients with lower SIR
Spleen/Muscle demonstrated a worse prognosis, possibly attributable to tumor-induced anemia. Schnitzbauer et- al discovered that preoperative anemia was an independent factor influencing CCA patients’ prognosis [
21].
In alignment with extant research, tumor size and lesion location are significantly correlated with the overall survival (OS) of patients diagnosed with ECC [
24]. Notably, individuals afflicted by pCCA tend to experience a more unfavorable prognosis compared to those affected by dCCA. This disparity can be attributed to the anatomical positioning of pCCA tumors around critical blood vessels and bile duct structures, rendering them more prone to infiltrating adjacent tissues and necessitating intricate hepato-biliary resection procedures [
25].
In addition, this study found that indicators such as GGT, ALT, and DBIL can reflect biliary obstruction and liver damage conditions, which are closely related to the prognosis of ECC patients. According to Zhang et al., high serum bilirubin concentration affects various pathological physiological metabolic processes including decreased albumin synthesis, reduced immune function, and altered hemodynamics [
26]. Complementarily, another study established an association between lower direct bilirubin levels and extended survival durations, corroborating the findings of our research [
27].
In our nomogram, we have incorporated eight variables, including gender, ALT, GGT, DBIL, tumor size, tumor location, SIRliver/muscle and SIRspleen/muscle. This convenient and practical tool enables clinical physicians to easily perform preoperative prognosis prediction through simple measurements for effective treatment decision-making. Furthermore, the accuracy and stability of the nomogram predictions were assessed through internal validation as well as 5-fold and 10-fold cross-validation in this study. The findings demonstrate that the nomogram exhibits robust stability.
This study, while insightful, is subject to certain limitations. Firstly, being a retrospective study, it is susceptible to selection bias. Secondly, the limited lesion size, MRI scan thickness, and resolution have resulted in infrequent collection of MR features of the lesions. Therefore, for future investigations, we intend to employ thin-layer MRI scans to explore the correlation between tumor characteristics and prognosis. Additionally, given the widespread integration of artificial intelligence in the medical domain, Radiomics emerges as a promising technology with immense potential for investigating the intricate relationship between radiological texture features and clinical outcomes, alongside molecular characteristics [
28]. Radiomics, by extracting comprehensive quantitative data that include both visible and subvisual elements from medical images, offers a more detailed and valuable perspective than traditional visual assessments by physicians [
29]. Deep learning (DL) networks, in contrast to conventional manual segmentation methods, provide accurate, objective automatic segmentation, mitigating errors and limitations inherent in manual processes [
30]. This not only broadens the range of information sources relied upon by physicians for patient condition assessment beyond a singular perspective but also effectively advances the practice of precision medicine [
31]. Given the aforementioned circumstances, future research will focus on expanding the sample size, conducting multicenter studies, and applying advanced deep learning techniques to improve our model’s predictive power. We also plan to integrate multidimensional data fusion methods to amalgamate information from various domains, enabling a comprehensive evaluation of patient risks. Additionally, close collaboration with expert medical teams will be sought to provide guidance for model design and optimization.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.