Advances in radiomics applications for intrahepatic cholangiocarcinoma: a comprehensive review and future directions

Image acquisition

High-quality and standardized images are typically obtained through imaging modalities such as MRI, CT, positron emission tomography - computed tomography(PET-CT), and ultrasound. In the clinical management of nearly all patients with ICC, CT or MRI imaging is routinely performed to acquire pertinent diagnostic images. Ensuring the quality and standardization of these images is crucial.

Region of interest segmentation

In most cases, experienced radiologists or radiation oncologists manually delineate the tumors. However, there has been development in automatic and semi-automatic segmentation methods, with some automatic approaches achieving results comparable to those of manual segmentation by radiologists. In the field of medical imaging, a variety of techniques are available for both semi-automatic and fully automatic segmentation, each employing tools of different levels of sophistication. Due to the inherently hypovascular nature of ICC, the tumor often exhibits distinct contrast with the surrounding normal liver tissue at certain stages, making the delineation of its boundaries relatively straightforward. Consequently, the performance of automated segmentation methods can even surpass that of manual tumor delineation by experienced radiologists. Basic segmentation methods primarily analyze the intensity and contrast within images, with a popular method being segmentation based on predefined thresholds. Alternatively, region-based segmentation identifies groups of similar and connected voxels using criteria based on uniformity^[15]. These methods are typically semi-automatic because they require human intervention to set thresholds or validate the uniformity criteria.

Further, some segmentation research incorporates models of uncertainty and algorithms for optimization to pinpoint specific areas within an image. Techniques such as statistical pattern recognition, which automatically identifies and categorizes patterns in images, c-clustering, which sorts images into distinct groups without supervision, and graph-based segmentation, which visualizes images as networks of voxels connected by edges optimized through a cost function, are frequently applied in the study of radiomics^[16,17].

Extraction and analysis of radiomics features

Multi-phase contrast-enhanced computed tomography (CE-CT) is regarded as the standard imaging modality for diagnosing ICC. ICC typically manifests as peripheral arterial phase hyperenhancement (APHE), peripheral washout, and delayed central enhancement. Consequently, the arterial phase of CT is considered the optimal timing for obtaining imaging data.

Radiomics features are divided into two categories: semantic features and agnostic features^[18]. Semantic features involve visible characteristics such as the size, shape, location, and degree of necrosis of tumor lesions. Although the analysis of these features is relatively easy to implement in clinical practice, they typically rely on the diagnostician’s experience, which can lead to variability in assessments both between different observers and within the same observer over time. In contrast, agnostic features are quantitative descriptors extracted from tissues of interest using mathematical methods and are not included in standard radiological reports. Agnostic features utilize advanced mathematical algorithms and are divided into morphological and statistical features. Morphological features describe the shape and physical structure of the segmented volume. Statistical features are further subdivided into first-order, second-order, and higher-order features. First-order features focus on the attributes of individual voxels, such as brightness or intensity, and include average intensity, maximum and minimum intensity, standard deviation, and histogram analysis, revealing the distribution and variation of pixel intensities within the image^[19]. Second-order features, also known as texture features, describe the texture patterns of an image by analyzing the spatial relationships between voxels. Commonly used second-order features include the gray-level co-occurrence matrix (GLCM), gray-level run-length matrix (GLRLM), and gray-level size zone matrix (GLSZM), which explore the spatial correlations of voxel intensities. Higher-order features employ more complex mathematical models and algorithms, such as wavelet transforms or deep learning technologies, to analyze the intricate relationships between voxels and extract deeper layers of image data. In ICC, morphological and statistical features are crucial for identifying texture patterns and heterogeneity within tumors. Features like the GLCM and wavelet transforms are particularly useful in discerning subtle differences that may indicate aggressive tumor behavior.

Predictive model building

Radiomics modeling primarily encompasses feature selection, modeling methodologies, and validation^[20]. This intricate process initiates with the extraction of pertinent quantitative features from a plethora of imaging attributes, encompassing aspects such as morphology, texture, and statistical properties. During the feature selection phase, sophisticated statistical analyses and advanced machine learning algorithms are employed to discern the most diagnostic or prognostic features. Commonly utilized techniques include principal component analysis (PCA), least absolute shrinkage and selection operator (LASSO) regression, and Random Forests, all of which are vital for mitigating feature redundancy and augmenting model efficacy.

Subsequently, in the modeling methodologies phase, the curated features are leveraged to construct predictive or classification models. Frequently employed modeling techniques encompass linear regression, logistic regression, support vector machines (SVM), Random Forests, and neural networks. Each technique possesses unique merits and specific applications. For instance, linear regression is apt for predicting continuous variables, logistic regression is utilized for binary classification challenges, whereas SVMs and Random Forests excel in managing high-dimensional data and non-linear problems. Neural networks, particularly deep learning models such as Convolutional Neural Networks, are remarkably proficient in processing intricate imaging data by autonomously extracting and amalgamating features, thereby enhancing model precision.

Finally, during the validation phase, the model’s performance and reliability are meticulously evaluated through both internal and external validation. Internal validation commonly adopts methods such as cross-validation and leave-one-out by partitioning the dataset into multiple subsets, sequentially utilizing one subset as the validation set while the remainder as the training set, to comprehensively assess the model’s performance across varying subsets. External validation employs an independent dataset to rigorously test the model, appraising its generalization capability and practical application value. Common performance metrics include accuracy, sensitivity, specificity, and the area under the ROC curve (AUC-ROC).

By adhering to these rigorous steps, radiomics is capable of extracting a copious amount of valuable quantitative data from medical images, unveiling potential correlations between imaging features and tumor heterogeneity. This capability not only facilitates a more precise assessment of the biological behavior and prognosis of tumors but also provides robust support for clinicians in diagnostic and therapeutic decision-making, thereby propelling the advancement of personalized medicine. By integrating radiomic features with clinical, pathological, and genetic data to construct multimodal predictive models, radiomics is progressively becoming an indispensable tool for enhancing clinical practice^[21-23][Figure 1].

Figure 1. The radiomics workflow.

Differentiating ICC from other diseases

The preoperative diagnosis of ICC currently relies mainly on imaging studies and CA19-9 levels^[7,24]. However, conventional imaging often provides limited information, and some patients’ CA19-9 levels remain within the normal range, making it challenging to accurately differentiate ICC from other liver lesions^[25]. Enhanced imaging offers some diagnostic differentiation value, with ICC typically showing progressive, centripetal enhancement, which can be distinguished from certain liver lesions. Nevertheless, observation of enhanced imaging is subjective, and due to the heterogeneity of ICC, some ICC imaging presentations are atypical and can overlap with other lesions, necessitating a precise and objective method to overcome these shortcomings. Radiomics, capable of mining and extracting information on intratumoral heterogeneity at high throughput, holds promise for differentiating ICC from intrahepatic bile duct stones^[26], other primary liver cancers^[27-33], etc.^[34,35].

In one study, Wang et al. delineated tumor region of interest (ROIs) and expanded them to create multiple ROI expansion areas (-2, 0, 2, 4, 6, and 8 mm) in MRI images of 87 hepatocellular carcinoma (HCC) patients and 75 pathologically confirmed ICC patients, establishing several predictive models^[36]. By combining convolutional features, the model’s performance was further improved. Compared to the standard ROI, expanded areas achieved better predictive performance, with the 6 mm expansion area exhibiting the best prediction ability [area under the curve (AUC): 0.84 vs. 0.86 vs. 0.88 vs. 0.90 vs. 0.94 vs. 0.82]. To differentiate HCC and ICC, Huang et al. developed a deep learning model based on channel and spatial attention mechanisms, called CSAM-Net, outperforming traditional radiomics models (RMs) constructed with logistic regression, LASSO regression, SVM, and random forests (AUC = 0.987 vs. 0.913)^[37]. Another study by Liu et al. indicated that using machine learning to analyze MRI and CT radiomics features for differentiating cHCC-ICC from HCC and ICC showed good predictive performance, potentially impacting treatment decisions^[28]. Additionally, Jiang et al. established a predictive model composed of 2 PET and 1 CT radiomics feature, finding that PET and CT radiomics features could effectively differentiate HCC and ICC, with an AUC of 0.86; clinical factors did not significantly enhance discrimination, with clinical models and combined models of radiomics features and clinical features yielding AUCs of 0.56 and 0.80, respectively^[38]. These studies demonstrate that radiomics can provide molecular-based imaging features and intratumoral heterogeneity information. When combined with clinical information, it can effectively differentiate HCC and ICC, suggesting that future radiomics should integrate various clinical factors to more effectively distinguish ICC from HCC. Differentiation between ICC and CHC is also a significant direction in clinical practice. Zhou et al. developed radiomics features based on MRI images, selecting 11 radiomics features and integrating alpha-fetoprotein, background liver disease (cirrhosis or chronic hepatitis) to construct a nomogram, showing strong calibration and discrimination performance, with training and validation cohort AUCs of 0.945 and 0.897, respectively^[39]. The radiomics nomogram proved superior to radiomics features and individual clinical models (P < 0.05). Xue et al., based on arterial phase CT images of 131 patients with intrahepatic bile duct stones and concurrent tumors, extracted four radiomics features, combined with three clinical features (fever, CA 19-9, and CEA), establishing a combined model^[26]. This model showed the best performance in predicting ICC concurrent with intrahepatic bile duct stones, with an AUC of 0.902. Another study by Xue et al. extracted radiomics features from arterial and venous phase enhanced CT images, building a model with phase-specific radiomics scores and two clinical factors (CEA and CA19-9), achieving the best model performance, with training and validation group AUCs of 0.864 and 0.843, respectively, surpassing simple RMs (training group AUC: 0.809, validation group AUC: 0.790) and clinical variable models (training group AUC: 0.801, validation group AUC: 0.830), indicating the combined model can distinguish ICC from inflammatory tumors with intrahepatic bile duct stones, improving diagnostic accuracy^[40]. Xu et al. utilized enhanced CT to extract texture features, identifying 38 eligible texture features^[32]. Through five feature selection methods (distance correlation, random forest, LASSO, etc.) and nine feature classification methods (linear discriminant analysis, SVM, logistic regression, etc.), 45 predictive models were established to differentiate ICC from liver lymphoma. Most models performed well (AUC > 0.85), with the random forest linear discriminant analysis showing the best performance among the 45 models (AUC = 0.997), suggesting that combining enhanced CT texture features with multiple machine learning models can effectively distinguish ICC from liver lymphoma. Xu et al., based on multiparametric MRI images, extracted radiomics features using the LASSO algorithm and selected clinical variables (tumor diameter and CEA level) and MRI results to construct a clinical model^[34]. The final radiomics nomogram model, which integrates both the clinical and RMs, demonstrated superior performance in differentiating intrahepatic mass-forming cholangiocarcinoma (IMCC) from colorectal liver metastasis (CRLM). The model achieved AUCs of 0.94 (95%CI: 0.90-0.97) for the training cohort, 0.93 (95%CI: 0.86-1.00) for the internal validation cohort, and 0.92 (95%CI: 0.84-1.00) for the external validation cohort. Table 1 summarizes the models and results of the analyzed studies.

Predicting LNM

LNM is significantly associated with poor prognosis in patients with ICC and is a crucial determinant of the tumor’s resectability^[24,42]. However, current methods for assessing lymph node status offer unstable predictive accuracy and limited discriminative capability^[43]. As a result, some patients initially deemed operable during pre-surgical assessments are later found to have metastasis during surgery, leading to abortive operations. Fine-needle aspiration, despite being invasive, has limited effectiveness in detecting minor lymph node metastases and carries a risk of tumor spread^[44,45]. The data from the papers are summarized in Table 2.

Radiomics prediction models built using features extracted from CT or MRI can be used to preoperatively predict LNM, thereby enabling clinicians to make better-informed clinical decisions^[49,50]. In 2019, Ji et al. extracted eight features related to lymph node status from CT images and, together with CA19-9 levels, built a combined predictive model^[48]. The model, tested on a training set of 103 patients and a validation set of 52 patients, showed AUCs of 0.8462 and 0.8921, respectively. It categorized LNM into high-risk and low-risk groups, revealing significant differences in overall survival (OS) and recurrence-free survival between the two groups. However, this study’s patient groups, both for training and validation, were from a single center and lacked external validation. In another study from 2019, Xu et al. constructed a SVM model based on five selected imaging features from MRI, creating a combined nomogram model with SVM scores, CA19-9 levels, and MRI characteristics^[50]. Compared to the standalone SVM model, the combined nomogram demonstrated superior discriminative ability in distinguishing between predicting LNM and non-LNM patients (AUC: training set: 0.842 vs. 0.788; validation set: 0.870 vs. 0.787). Zhang et al. integrated radiomics features from multiple CT sequences to build a fusion model, which, across three cohorts (training, external validation, and internal validation), showed AUC values surpassing those of clinical RMs^[45]. The constructed nomogram, validated independently, displayed good differentiation and prognostic value. In summary, employing radiomics prediction models based on features extracted from CT or MRI holds promise for accurately predicting preoperative lymph node status, potentially becoming a valuable tool for preoperative assessment and prognostic evaluation in ICC.

Predicting microvascular invasion

Microvascular invasion is an independent risk factor affecting the prognosis of patients undergoing radical resection for ICC^[42,51], with those experiencing predicting microvascular invasion (MVI) often facing a poorer prognosis^[52]. However, clinical detection of MVI is primarily conducted through microscopic histopathological examination, identifiable only postoperatively. Therefore, building models to predict MVI preoperatively, thus offering references for clinicians to adjust surgical methods and extents, has emerged as a hot topic in ICC research recently. The advent and development of radiomics, by enabling the extraction of features from the entire tumor and its surrounding areas, have continually improved the accuracy of preoperative MVI predictions. Table 3 summarizes the models and results of the analyzed studies.

In recent years, numerous studies utilizing radiomics to predict preoperative MVI have surfaced, trying various imaging modalities including ultrasound, CT, MRI, and PET-CT for model construction^[38,53-56]. Qian et al. utilized the LASSO logistic regression to identify the optimal features from MRI images of patients in three cohorts (training cohort: n = 130, validation cohort: n = 33, test cohort: n = 24) and developed a radiomics model for preoperative MVI prediction^[57]. This model showed excellent and consistent predictive performance in the training (AUC = 0.950), validation (AUC = 0.883), and test (AUC = 0.812) cohorts. Additionally, Qian et al. employed univariate and multivariate analyses to identify independent predictors of MVI status: larger tumor size (P = 0.003) and intrahepatic bile duct dilatation (P = 0.002)^[57]. These factors were combined with the final radiomics model to create an MVI prediction nomogram, which also achieved strong predictive results across the training (AUC = 0.953), validation (AUC = 0.861), and test (AUC = 0.819) cohorts. In 2023, Ma similarly built a predictive model for MVI using MRI features. Unlike Qian et al.^[57], Ma et al. compared the predictive capabilities of different MRI sequences and VOIs, selecting the three most optimal sequences (T1WI-D, T1WI, DWI) and one optimal VOI (including the tumor and a 10 mm peritumoral area), achieving desirable prediction outcomes (AUCTC = 0.987 and AUCVC = 0.859)^[56]. The predictive ability further improved when combining tumor size, intrahepatic bile duct dilatation, and a multi-sequence fusion of VOI10 mm MRI in a radiomics model: TC (AUC = 0.995, 95%CI: 0.987-1.000) and VC (AUC = 0.867, 95%CI: 0.798-0.921). Additionally, Zhou et al. and Chen et al. developed RMs based on dynamic contrast-enhanced MRI images, achieving satisfactory results^[54,55]. Beyond MRI, other imaging modalities also exhibited good predictive capabilities. In 2021, Xiang et al. constructed a portal phase image radiomics model with six features extracted from enhanced CT, displaying good predictive power in two cohorts (training group: n = 110, AUC = 0.804; validation group: n = 47, AUC = 0.769)^[53]. The integration of significant clinical factors {satellite nodules [odds ratio (OR)=13.73], arterial hypointensity (OR = 4.31), and tumor contour (OR = 4.99)} with radiomics features into a nomogram achieved satisfactory predictive results, with AUCs of 0.886 in the training group and 0.80 in the validation group, significantly surpassing the simple radiomics model and the clinical model (training group AUC: 0.822, test group AUC: 0.756). Fiz et al. elucidated that preoperative CT portal phase extraction of tumor and peritumoral tissue imaging characteristics enhanced the prediction of ICC grading and MVI (clinical data model: AUC = 0.75; clinical data model + tumor-VOI: AUC = 0.82; clinical data model + tumor-/edge-VOI: AUC = 0.82), finding that combining tumor with peritumoral tissue imaging characteristics could improve predictive power^[58]. Jiang et al., by extracting features from PET and CT to construct predictive models separately, showed that the PET group significantly outperformed the CT group in predicting MVI (PET: AUC = 0.88; CT: AUC = 0.67)^[38]. However, the predictive ability of models combining PET and CT features decreased (PET + CT: AUC = 0.75). The best predictive results were achieved with a nomogram combining PET, CT, and clinical factors (PET + CT + Clinical: AUC = 0.90). It is evident that preoperative imaging-based prediction of MVI holds value, aiding doctors in making clinical decisions related to treatment plans and influencing.

Early recurrence post-surgery

Early recurrence is defined as recurrence within two years post-surgery or within one year postoperatively. It is a critical factor affecting overall patient survival^[59,60]. Therefore, effectively identifying patients at high risk of early recurrence can help clinicians devise targeted treatment plans, thereby enhancing patient survival time. While some prediction models based on clinical features exist, their accuracy and applicability are limited, leading to an urgent need for new methods to identify ER patients. Medical imaging plays a crucial role in the preoperative assessment of ICC^[61], making models built on image features for predicting ICC’s early recurrence a focus of research^[62-66].

In 2018, Liang et al. were the first to use nine features extracted from MRI images to build a radiomics model to predict patients with ICC prone to early recurrence^[67]. In the training cohort (n = 139) and the validation cohort (n = 70), the AUC of radiomics features was 0.82 (95%CI: 0.74-0.88) and 0.77 (95%CI: 0.65-0.86), respectively. Liang et al. then developed a radiomics nomogram that combined imaging features and tumor clinical staging, achieving better predictive capacity with AUCs of 0.90 (95%CI: 0.83-0.94) and 0.86 (95%CI: 0.76-0.93) across both cohorts^[67]. Differing from Liang et al.^[67], Xu et al. not only focused on the internal structure features of tumors in MRI images but also integrated peritumoral area features (3 and 5 mm), finding that models combining internal tumor structure and a 5 mm peritumoral area performed better than those that only considered internal tumor structure or combined it with a 3 mm peritumoral area (AUC: 0.852 vs. 0.835 vs. 0.760)^[47].

Beyond MRI, enhanced CT has also been frequently used to construct models predicting postoperative recurrence. Studies by Song et al. utilizing CT-constructed radiomics prediction models for early recurrence optimized the model’s predictive capacity by integrating clinical features. Song et al.’s combined prediction model, incorporating 15 radiomics features and three clinical features (CA19-9 > 1,000 U/ml, vascular invasion, and tumor margin), achieved an AUC of 0.974, significantly outperforming standalone RMs (AUC = 0.877), clinical feature models (AUC = 0.733), and the AJCC 8th TNM staging system (AUC = 0.717)^[66]. The studies by Bo et al. and Xu et al. also demonstrated that a combined model integrating radiomics and clinical features surpassed individual models (AUC: 0.873 vs. 0.872 vs. 0.685, 0.85 vs. 0.82 and 0.75)^[46,68]. Thus, selecting and combining different models appropriately can potentially enhance the new model’s predictive capacity. Yang et al. developed three predictive models, composed of clinical factors (CA19-9), pathological factors (MVI, tumor differentiation grade), imaging features, and a combination of these factors^[69]. The composite model, integrating clinical factors (CA19-9), pathological factors (MVI, tumor differentiation grade), and imaging features, achieved the best predictive capacity (AUC = 0.876 vs. 0.823 vs. 0.697). In another study, Zhao et al. established three sets of models for prediction, similarly to Yang, but incorporated postoperative immunohistochemical staining of tissue specimens for epidermal growth factor receptor (EGFR), vascular endothelial growth factor receptor (VEGFR), P53, and Ki67 into the pathological factors^[65]. The results, akin to Yang’s et al. study^[69], showcased the superiority of the composite model, which, compared to standalone radiomics and CRP models, demonstrated enhanced ER predictive performance with an AUC, sensitivity, and specificity of 0.949, 0.875, and 0.774, respectively. Observing the outcomes of these studies reveals that MRI and CT can effectively predict postoperative recurrence, and appropriately adding factors like clinical features, imaging features, postoperative pathology results, and immunological factors can significantly improve the predictive ability of the model. Studies are listed in Table 4.

OS

The most commonly used staging system for ICC currently is the TNM staging, based on the tumor, lymph nodes, and metastasis. Due to tumor heterogeneity, the prognosis of ICC patients varies significantly among individuals, making the prediction accuracy somewhat limited^[70,71]. In clinical practice and future clinical trials, effective prognostic prediction tools, such as radiomics-based predictive models and nomograms, have been proposed^{[35,61,72-74]}. This represents a pressing need for formulating treatment strategies.

Deng et al. conducted a radiomics analysis of venous and arterial phase CT images of 82 ICC patients, combining radiomics features with clinical factors to establish a composite model^[75]. This model included the psoas muscle index, radiomics score, intrahepatic bile duct stones, carcinoembryonic antigen, and the neutrophil-to-lymphocyte ratio as five indicators. The results showed a C-index of 0.768, with the AUC for predicting 1-year and 3-year OS rates being 0.809 and 0.886, respectively, significantly higher than other models. Park et al. aimed to develop and validate a preoperative model that could predict postoperative outcomes^[76]. They constructed three different models using clinical, radiological, and radiomics features, with the focus on predicting recurrence-free survival. The clinical-radiological-radiomics model showed the best performance, with a training group C-index of 0.75 (0.72-0.79). This assists in assessing the postoperative outcomes of ICC patients with tumor formations preoperatively, allowing for the selection of the best treatment plan at the initial decision-making stage.

Fiz et al. extracted radiomics features from 18F-FDG PET/CT images and, along with clinical data, established clinical models, clinical plus intratumoral RMs, and clinical plus intratumoral and peritumoral (5 mm) RMs to predict tumor grading, MVI, OS, and progression-free survival (PFS)^[35]. The results indicated that adding peritumoral radiomics features could optimize predictions for ICC tumor grading and survival rates (AUC = 0.834 vs. 0.783 vs. 0.718), but did not improve MVI predictions (AUC = 0.881 vs. 0.871 vs. 0.773). Thus, ICC radiomics based on PET can also predict pathological data, allowing for a reliable preoperative prognostic assessment. Tang et al. conducted a radiomics analysis of CT images and used the LASSO method to select three imaging features^[77]. Multivariate Cox analysis identified three independent prognostic factors: cirrhosis, CA19-9 levels ≥ 35 U/mL, and tumor size > 5 cm. A nomogram combining imaging features and clinical factors was constructed. The radiomics nomogram demonstrated significant prognostic value for OS. Tang et al.’s study revealed notable differences in 1-year and 3-year survival rates between high-risk and low-risk patients: 30.4% vs. 56.4% and 13.0% vs. 30.6%, respectively (P = 0.018). Additionally, ultrasound could also be utilized to construct models^[77]. Li et al. were the first to extract radiomics features from baseline ultrasound (US) and contrast-enhanced ultrasound (CEUS) images (four sequences) to build a preoperative model predicting OS in patients with ICC^[73]. The nomogram, including CA 19-9, gender, ascites, radiomics features, and radiological characteristics, achieved a higher C-index than the 8th TNM staging system. In another study, Yang et al. used an MRI-based radiomics model to predict preoperative survival outcomes^[78]. Integrating radiomics features into the TNM staging system significantly improved prognostic accuracy (validation set C-index 0.745 vs. 0.649, P = 0.039, NRI improved by 39.9%-43.8%, IDI improved by 16.1%-19.4%). In summary, as a method of extracting quantitative and high-dimensional features from medical imaging data, radiomics can accurately describe tumor biological characteristics, predicting the prognosis of ICC. Studies are listed in Table 5.