Artificial intelligence in the diagnosis and prognosis of intrahepatic cholangiocarcinoma: Applications and challenges

Abstract

Intrahepatic cholangiocarcinoma (ICC) is a primary liver malignancy with increasing global incidence and mortality rates. The 5-year overall survival rate for patients with ICC is approximately 9%. Surgical resection currently represents the only curative treatment option. However, due to the high aggressiveness, insidious onset, and atypical clinical presentation of ICC, many patients either miss the optimal surgical window or experience early postoperative recurrence and metastasis. This poses significant challenges for hepatobiliary surgeons worldwide. Artificial intelligence (AI), as a prominent driver of technological advancement, offers promising new avenues for managing ICC. By leveraging powerful machine learning and deep learning algorithms, AI has demonstrated promising outcomes in ICC diagnosis, particularly in differentiating it from hepatocellular carcinoma, and in predicting critical prognostic factors such as early recurrence, lymph node metastasis, and microvascular invasion. These innovations can support clinical decision-making and ultimately improve patient outcomes. Future efforts should prioritize robust clinical studies evaluating the effectiveness of AI in ICC management.

Key Words: Artificial intelligence; Intrahepatic cholangiocarcinoma; Early diagnosis; Prognostic prediction; Machine learning

Core Tip: The emergence of artificial intelligence (AI) has introduced new possibilities for improving the poor prognosis associated with intrahepatic cholangiocarcinoma. This review summarizes recent advances in AI applications related to the diagnosis, differential diagnosis, and prediction of recurrence risk factors, early recurrence, survival, and treatment response in intrahepatic cholangiocarcinoma. Additionally, the numerous challenges associated with AI development are discussed, including nascent legal frameworks, vulnerabilities in data privacy protection, and a lack of empirical research in clinical settings.

INTRODUCTION

Intrahepatic cholangiocarcinoma (ICC) is the second most prevalent primary malignant liver tumor in humans, following hepatocellular carcinoma (HCC), and constitutes approximately 10%-15% of all primary liver cancers[1]. Recently, the global incidence and mortality rates have begun to increase[2]. ICC typically presents with an insidious onset, lacking specific clinical manifestations in its early stages, often leading to an asymptomatic presentation. As the disease progresses, symptoms such as weight loss, abdominal discomfort, jaundice, hepatomegaly, and palpable abdominal masses may emerge[3]. Consequently, most patients (approximately 70%) are diagnosed at an advanced stage, severely limiting treatment options and contributing to a poor prognosis[4,5].

Surgical resection remains the only potentially curative modality for ICC; however, due to late-stage diagnosis, patients often present with locally advanced (unresectable) or metastatic disease, rendering only approximately 25% eligible for resection[5,6]. In unresectable cases, palliative therapy is the sole option. The first-line chemotherapeutic regime currently recommended is combined gemcitabine and cisplatin (GemCis regimen), but this approach is not curative and yields a moderate overall survival (OS) of less than one year[7,8]. Furthermore, locoregional therapies, radiotherapy, immunotherapy, and targeted therapies have not demonstrated significant efficacy in improving survival outcomes[9-11]. Meanwhile, recurrence rates up to 50%-60% have been reported following surgical resection, with a median recurrence-free survival of 26 months[3,12]. The National Cancer Institute reported a 5-year OS rate of approximately 9% for patients with ICC[13], presenting a substantial challenge for clinicians and researchers, as well as imposing a considerable economic burden on healthcare systems worldwide.

Early diagnosis of ICC is hindered by the absence of well-defined risk factors, unlike HCC, which typically develops from liver cirrhosis. Consequently, ICC is typically detected incidentally during cross-sectional imaging for unrelated issues[14]. Imaging, such as ultrasound (US), computed tomography (CT), magnetic resonance imaging (MRI), and ¹⁸F-fluorodeoxyglucose (¹⁸F-FDG) positron emission tomography (PET), is crucial for diagnosis, staging, treatment planning, and monitoring of ICC. However, in clinical practice, conventional imaging techniques often struggle to differentiate ICC from HCC and metastatic liver lesions originating from extrahepatic primary tumors[15].

While unresectable ICC cases primarily respond to conventional chemotherapy, targeted therapy, and immunotherapy, HCC benefits from transarterial chemoembolization, targeted therapy, and immunotherapy[7,16]. Accordingly, accurately distinguishing ICC from HCC before initiating treatment is critical. Additionally, factors such as large tumor diameter, multiple lesions, vascular invasion, and lymph node (LN) metastasis (LNM) are key predictors of early recurrence (ER) after resection and OS[12,17]. Therefore, when selecting surgical candidates, it is crucial to incorporate these considerations, alongside treatment goals, life expectancy, and cost-benefit analyses, into a personalized surgical stratification approach to improve postoperative survival rates[12]. Among these, LNM is among the most significant prognostic factors; patients with LNM who undergo resection have a 5-year survival rate below 20%[3]. However, preoperative imaging using conventional methods fails to predict regional LNM reliably[18]. This highlights an urgent need for novel approaches to overcome the limitations of current techniques.

Artificial intelligence (AI) represents a specialized branch of computer science dedicated to developing machines capable of replicating aspects of human intelligence (Figure 1)[19]. AI applications have been investigated across diverse medical fields, including healthcare[20], drug discovery[21], and oncology[22]. Compared with traditional biometric approaches, AI offers greater flexibility and scalability, enabling its use in addressing complex tasks. Additionally, AI can effectively integrate vast amounts of heterogeneous data types and discern complex relationships among variables in a flexible and trainable manner. Machine learning, a key subfield of AI (Figure 2), employs algorithms that enable computers to learn from large datasets independently and to make predictions or decisions regarding tasks relevant to clinical practice or biomedical research[19]. Based on label availability, machine learning can be categorized into supervised, unsupervised, semi-supervised, reinforcement, and ensemble learning, with the latter combining multiple algorithms[19]. Deep learning is a subset of machine learning, with its concepts originating from research on artificial neural networks. Its technical strength lies in extracting more abstract features from large and complex datasets, such as images, text, and speech, through multi-layered processing[23].

Figure 1 Relationship between artificial intelligence, machine learning, and deep learning.

Figure 2 Workflow of machine learning and deep learning. ICC: Intrahepatic cholangiocarcinoma.

The concept of radiomics was introduced by Lambin et al[24] in 2012. Unlike traditional qualitative interpretations, radiomics uses high-throughput extraction of quantitative features from medical images to elucidate associations between imaging findings and disease states, thereby aiding diagnosis, treatment planning, and prognostication[24]. With the exponential expansion of medical data and advancements in computational capabilities, AI excels at efficiently processing large datasets. Within clinical practice, AI systems not only support clinicians in managing routine tasks but also offer opportunities for enhancing healthcare standards, particularly in under-resourced settings[25]. Specifically, for ICC, the progression of AI presents promising opportunities to improve diagnostic and therapeutic modalities. Hence, the European Union has pledged financial support for AI and machine learning research and development initiatives targeting ICC[5]. The present review focuses on recent advances and challenges of AI in the diagnosis and prognosis of ICC.

AI IN ICC DIAGNOSIS PROCESS

The subtle onset and diagnostic challenges of ICC often delay intervention. Early detection and diagnosis expand therapeutic options and improve prognosis[5,14]. As the two most prevalent primary liver cancers, ICC and HCC exhibit substantial differences in terms of management and outcomes[7,16]. However, distinguishing between them using current non-invasive techniques remains difficult[15]. AI models based on medical imaging (primarily US, CT, and MRI) have demonstrated excellent performance in diagnosing and differentiating ICC from other malignancies (Table 1). Furthermore, AI can analyze digitized histopathological slides and assist pathologists in enhancing diagnostic accuracy and efficiency. Computational approaches using high-throughput multi-omics analyses can identify novel biomarkers for effective ICC screening and diagnosis.

Table 1 Applications of artificial intelligence based on medical imaging for intrahepatic cholangiocarcinoma diagnosis.

Ref.	Sample size	Data source	Algorithms	Aim	Validation set or test set results
Ding et al[27]	3725 cases	CEUS	LSTM, MLP	ICC diagnosis	Accuracy (97%)
Ren et al[29]	226 cases	US	SVM	ICC vs HCC	AUC (0.936), sensitivity (90%), specificity (85.7%), accuracy (86.8%)
Chen et al[31]	465 cases	B-mode US	ResNet18	ICC vs HCC vs cHCC-CCA	AUC (0.9237), sensitivity (84.59%), specificity (92.65%), accuracy (86%)
Qian et al[32]	169 cases	2D US	LASSO, RFE, RF	ICC vs IBDS	AUC (0.988)
Gao et al[36]	723 cases	Multi-phase CECT	CNN, RNN	ICC vs HCC	AUC (0.944), accuracy (82.9%)
Wei et al[38]	4039 cases	Multi-phase CECT	ResNet50, SKD	ICC vs HCC vs metastatic liver cancer	AUC (0.956), accuracy (88.7%)
Midya et al[39]	814 cases	Portal venous phase CECT	Inception v3	ICC vs HCC vs CRLM vs liver benign tumors	Accuracy (96.27%)
Xue et al[40]	96 cases	Arterial-phase CECT	LASSO	IBDS with ICC vs IBDS with cholangitis	AUC (0.879)
Yang et al[42]	112 cases	CECT	LASSO	ICC vs EHA	AUC (0.868), sensitivity (94.4%), specificity (81.3%)
Xu et al[43]	129 cases	CECT	RF, LDA	ICC vs HL	AUC (0.997), accuracy (96.9%)
Liu et al[44]	177 cases	DCE-MRI	LASSO	ICC vs HCC	AUC (0.877)
Hu et al[46]	514 cases	Multi-phasic MRI	TPOT	ICC vs HCC	AUC (0.79), accuracy (75%), sensitivity (75%), specificity (79%)
Liu et al[47]	112 cases	T2-weighted MRI	SFFNet	MF-ICC vs HCC	AUC (0.968), accuracy (92.26%)
Zhou et al[48]	216 cases	DCE-MRI	LASSO	MF-ICC vs cHCC-CCA	AUC (0.897), sensitivity (79%), specificity (76.1%), accuracy (76.9%)
Xu et al[49]	133 cases	Multiparameter MRI	mRMR, LASSO	MF-ICC vs CRLM	AUC (0.94)
Starmans et al[50]	486 cases	T2-weighted MRI	WORC	ICC vs HCA vs FNH	AUC (0.78), sensitivity (84%), specificity (62%), accuracy (71%)
Cheng et al[51]	178 cases	CECT, MRI	ResNet50, LASSO	ICC diagnosis	AUC (0.937), sensitivity (80%), specificity (87.2%), accuracy (85.2%)
Jiang et al[52]	127 cases	18F-FDG PET/CT	SFFS, RF	ICC vs HCC	AUC (0.86), sensitivity (78%), specificity (88%), accuracy (82%)

CEUS: Contrast-enhanced ultrasound; US: Ultrasound; CECT: Contrast-enhanced computed tomography; DCE-MRI: Dynamic-contrast-enhanced magnetic resonance imaging; MRI: Magnetic resonance imaging; 18F-FDG PET/CT: 18F-fluorodeoxyglucose positron emission tomography/computed tomography; LSTM: Long short-term memory; MLP: Multilayer perceptron; SVM: Support vector machine; ResNet: Residual neural network; LASSO: Least absolute shrinkage and selection operator; RF: Random forest; RFE: Recursive feature elimination; CNN: Convolutional neural network; RNN: Recurrent neural network; SKD: Self-knowledge distillation module; LDA: Linear discriminant analysis; TPOT: Tree-based pipeline optimization tool; SFFNet: Strided feature fusion residual network; mRMR: Maximum relevance and minimum redundancy; WORC: Workflow for optimal radiomics classification; SFFS: Sequential forward floating selection; ICC: Intrahepatic cholangiocarcinoma; HCC: Hepatocellular carcinoma; cHCC-CCA: Combined hepatocellular-cholangiocarcinoma; IBDS: Intrahepatic bile duct stones; CRLM: Colorectal liver metastasis; EHA: Early hepatic abscess; HL: Hepatic lymphoma; MF-ICC: Mass-forming intrahepatic cholangiocarcinoma; FNH: Focal nodular hyperplasia; HCA: Hepatocellular adenoma; AUC: Area under the curve.

AI in medical image-based diagnosis

US: ICC most commonly manifests as a hypoechoic mass and segmental bile duct dilatation in non-contrasted US images[3,4]. Following contrast injection, washout occurs within 60 seconds in ICC. During the portal venous phase, ICC exhibits a more pronounced washout intensity than HCC[26]. Hence, contrast-enhanced US (CEUS) offers superior capability for ICC diagnosis compared to conventional US. To better leverage the potential of CEUS, Ding et al[27] developed a model using long short-term memory and multilayer perceptron networks to diagnose ICC based on CEUS videos collected from 52 centers. The model achieved a 97% accuracy in the test set, enabling junior radiologists to achieve diagnostic performance comparable to that of senior radiologists. However, CEUS interpretation is heavily reliant on operator experience, particularly in cases of chronic liver disease. Approximately half of patients with ICC demonstrate uniform high enhancement during the arterial phase, followed by washout in the venous phase, closely mimicking HCC and posing significant diagnostic challenges[28]. Ren et al[29] constructed a preoperative differential diagnosis model for ICC and HCC using US radiomic features with support vector machine (SVM), achieving an area under the curve (AUC) of 0.936.

Combined hepatocellular-cholangiocarcinoma (cHCC-CCA) is a rare form of liver cancer that exhibits hepatocellular and biliary differentiation[30], making it difficult to distinguish from HCC and ICC on imaging and pathology[7]. Chen et al[31] developed four deep learning models based on US images to differentiate ICC, HCC, and cHCC-CCA; residual neural network 18 offered the highest accuracy and robustness, with an AUC of 0.9237 in the independent test cohort. However, the US systems and scanning parameters used in the study were heterogeneous, requiring external validation by other institutions to assess generalizability. Intrahepatic bile duct stones (IBDS), a risk factor for ICC development[1,5], can be difficult to differentiate preoperatively from ICC. Qian et al[32] developed a robust and reliable diagnostic tool for preoperative differentiation between these two conditions using a combined model that integrated clinical variables and US radiomic features selected via the least absolute shrinkage and selection operator (LASSO) and recursive feature elimination, achieving an AUC of 0.988.

CT: CT is considered the standard imaging modality for the preoperative characterization and staging of ICC, enabling a comprehensive assessment of the primary tumor’s relationship with adjacent structures and potential thoracic/abdominal metastases[33]. The typical CT appearance of ICC is a non-enhancing hypodense liver mass with irregular margins, demonstrating peripheral rim enhancement during the arterial phase and progressive hyperattenuation in the portal venous and delayed phases[3,4]. This enhancement pattern correlates with the tumor’s histological architecture; increased perfusion induced by cancer cells concentrated in the periphery accounts for peripheral arterial enhancement, whereas central hypovascularity and progressive enhancement reflect a central tumor core characterized by sparse tumor cells and a proliferative fibrous stroma[34]. In contrast, HCC typically exhibits rapid arterial enhancement followed by hypoattenuation in the venous or delayed phase[3]. In clinical practice, this imaging discrepancy generally facilitates the effective differentiation between ICC and HCC. Unfortunately, ICCs that exhibit arterial nodular hyperenhancement are increasingly encountered, particularly in patients with chronic liver disease or cirrhosis[35]. This atypical ICC enhancement pattern mimics HCC, potentially leading to misdiagnosis and impacting clinical management. To address this, Gao et al[36] developed an effective differential diagnosis model by combining deep convolutional neural networks (CNNs) and gated recurrent neural networks to extract spatial and temporal features from multi-phase CECT images, achieving an 82.9% accuracy in distinguishing ICC from HCC.

Additionally, the unique structure of the liver provides a favorable environment for cancer cell survival, making it one of the organs most susceptible to metastasis[37]. The lack of standardized imaging features for diagnosing secondary liver metastases often results in heavy reliance on the radiologist’s experience and PET reports. However, the former is highly subjective, whereas the latter is too expensive for widespread use in resource-limited regions. Wei et al[38] addressed this with liver lesion network (LiLNet), a residual neural network 50-based network, achieving an 88.7% accuracy in distinguishing ICC, HCC, and metastatic liver cancer. LiLNet was trained on data from 2888 patients across two large medical centers in China and externally validated using data from 1151 patients at four additional domestic centers. The substantial dataset size underpins the generalizability of their diagnostic model. Similarly, Midya et al[39] applied a modified inception v3 network to classify HCC, ICC, colorectal liver metastasis (CRLM), the most common type of liver metastasis, and benign tumors, reaching an overall classification accuracy of 96.27%, with a specific ICC accuracy of 95.83%. However, although the data were collected from multiple centers, the retrospective design of the study by Midya et al[39] may have introduced selection bias in patient enrollment. Thus, given that the tumors of patients who received resection were smaller and confined to a single lobe, further prospective validation is warranted.

In addition to differentiating malignancies, efficient exclusion of benign lesions is crucial for ICC diagnosis. Xue et al[40] combined an arterial-phase CT radiomic model with three clinical features (fever, carcinoembryonic antigen, and carbohydrate antigen 19-9 (CA 19-9) into a combined model to differentiate between IBDS complicated by ICC and simple IBDS with cholangitis. In the external validation cohort, this combined model (AUC = 0.879) outperformed the standalone radiomic model (AUC = 0.824) and the clinical feature model (AUC = 0.755). Hepatic abscess is a common hepatic lesion; however, its early-stage imaging manifestations are often non-specific, posing challenges in the differentiation from neoplastic lesions[41]. To differentiate between early hepatic abscesses from ICC preoperatively, Yang et al[42] constructed a nomogram using a radiomic score and clinical features, achieving an AUC of 0.868 in the validation cohort. Xu et al[43] used a random forest (RF) model to distinguish ICC from hepatic lymphoma, achieving an AUC of 0.997 and accuracy of 96.9%.

MRI and PET: MRI may offer advantages over CT for diagnosing and staging ICC[7]. On MRI, ICC typically appears hypointense on T1-weighted images and hyperintense on T2-weighted images. T2-weighted images may also show central hypointensity corresponding to the fibrotic areas[4]. Although diffusion-weighted imaging is an effective tool for detecting malignancies, it has limited use in differentiating ICC from HCC[3]. The advent of AI offers new opportunities for differentiating ICC from HCC using MRI.

Liu et al[44] employed LASSO to select optimal radiomic features from dynamic contrast-enhanced (DCE) MRI data to build a model, achieving an AUC of 0.877 in the validation cohort. However, the development and manual optimization of radiomic models require substantial machine learning expertise, potentially hindering clinician adoption[19]. Automated machine learning automates key, time-consuming, complex, and expertise-dependent steps within the workflow, lowering the barriers to machine learning applications while enhancing scientists’ productivity[45]. Hu et al[46] applied manual and automated analyses on multiphasic MRI data to determine the optimal machine learning model for differentiating the two primary liver cancers. They found that the performance of automated machine learning was comparable to that of manual optimization, achieving sensitivity and specificity similar to radiologists in classifying ICC and HCC. Most radiomic studies focus on intratumoral features, while overlooking the peritumoral region, which provides significant diagnostic information for hepatic cancers[33]. To capture more peritumoral features, Liu et al[47] proposed a “semi-segmentation” preprocessing method for defining regions of interest on T2-weighted images of mass-forming ICC (MF-ICC) and HCC. Their model achieved an overall accuracy of 92.26% and AUC of 0.968.

In addition to distinguishing HCC, there have been other AI applications for ICC diagnosis using MRI. Zhou et al[48] extracted radiomic features from DCE-MRI and combined them with alpha-fetoprotein levels and history of other liver diseases (cirrhosis or chronic hepatitis) to construct a nomogram for differentiating MF-ICC from cHCC-CCA. This achieved an AUC of 0.897 in the validation cohort. AI can also effectively differentiate between metastatic and benign liver lesions from ICC. Xu et al[49] used maximum relevance and minimum redundancy and LASSO features to construct a clinical-radiomic nomogram to distinguish MF-ICC from CRLM, achieving an AUC of 0.94. Starmans et al[50] used the workflow for an optimal radiomics classification toolbox, which integrates eight machine learning algorithms, to develop a radiomics model for differentiating ICC from benign lesions (hepatic adenoma and focal nodular hyperplasia). The model showed consistent performance between the internal validation (AUC = 0.78) and external validation (AUC = 0.76). ICC exhibits distinct features on CT and MRI, and clinicians typically need to integrate results from both modalities for comprehensive assessment. However, most radiomic studies rely on either CT or MRI alone, which can potentially lead to inconsistent conclusions among single-modality models. Cheng et al[51] innovatively fused CT and MRI radiomic features to build a combined model to distinguish ICC from other primary liver cancers. This fused model achieved the highest AUCs in both the training (0.994) and test (0.937) cohorts compared with CT-only or MRI-only models.

For patients deemed to have potentially resectable disease based on CT or MRI, ¹⁸F-FDG-PET/CT is effective in identifying distant and LN metastases, enabling accurate tumor staging and appropriate surgical candidate selection[33]. However, traditional imaging analysis for evaluating the primary ICC lesion lacks the advantages of ¹⁸F-FDG-PET/CT[3]. Jiang et al[52] applied sequential forward floating selection and the RF algorithm to build a supervised machine learning classifier for the preoperative identification of ICC and HCC using PET/CT data, achieving an AUC of 0.86 in the test cohort.

AI-assisted pathological examination

The most common pathological type of ICC is adenocarcinoma, characterized by tubular or papillary structures and a variable fibrous stroma[53,54]. Although clinical presentation, laboratory tests, and imaging assessments provide valuable clues for ICC diagnosis, histopathological examination remains the gold standard for a definitive diagnosis[3,7]. AI is increasingly used to assist in histopathological diagnosis across different organ systems by using digitized whole-slide images (WSIs) stained with hematoxylin and eosin. This not only streamlines workflows but also helps pathologists tackle diagnostically challenging cases[22]. Here, AI applications for the pathological diagnosis of ICC are summarized.

Kiani et al[55] developed an AI-assisted diagnostic tool using CNNs to aid pathologists in differentiating ICC from HCC on WSIs and investigated its impact on diagnostic accuracy. A subset of pathologists significantly improved their diagnostic accuracy after using this tool (P = 0.045). The “black box” effect represents a significant challenge in deep learning. As deep learning has achieved notable success across the medical field, applications increasingly rely on CNNs and deep neural networks. However, the complexity and lack of interpretability inherent in deep learning models hinder our understanding of their internal decision-making processes, a phenomenon known as the “black box” effect[23]. To address this challenge, Li et al[56] used heatmap visualization techniques to elucidate the features extracted by their model across different types of liver cancers. By providing interpretability for the diagnostic process, these visualizations enhance clinicians’ trust in and understanding of the model’s function. As previously mentioned, the complex histological composition of cHCC-CCA often makes it difficult to distinguish from ICC, leading to misdiagnosis and posing significant diagnostic challenges for pathologists[30]. Calderaro et al[57] analyzed 405 patients who were initially diagnosed with cHCC-CCA. Using a supervised deep learning model constructed by attention-based multiple instance learning, the tumors were reclassified as either ICC or HCC. This reclassification aligns with clinical outcomes, genetic variations, and spatial in situ gene expression profiles.

Pathological diagnosis of liver adenocarcinomas is common but complex. ICC and CRLM are the most common primary and secondary hepatic adenocarcinomas, respectively. However, ICC typically has a poorer prognosis than CRLM due to its aggressive nature[39], and treatment approaches differ significantly. Consequently, clinical decision-making heavily relies on accurate pathological diagnosis. Albrecht et al[25] developed a pathological diagnostic tool based on EfficientNet B3, which, in a validation study involving 50 patients, outperformed six pathologists in distinguishing between ICC and CRLM tumors.

Applications of AI in identifying and characterizing biomarkers

Biomarkers are key in diagnosing, staging, and treating ICC[4,5]. CA 19-9 is a commonly used biomarker for ICC diagnosis but has limited sensitivity (62%) and specificity (63%), particularly for early-stage detection[3]. AI-driven omics analyses may help identify and characterize promising biomarkers, potentially revolutionizing cancer diagnostics. Kajornsrichon et al[58] identified three autoantibody biomarkers in 26 serum samples that, when combined with CA 19-9, improved ICC diagnostic accuracy (AUC = 0.838) compared with CA 19-9 alone. However, the small sample size represents a notable limitation of their study. Zhang et al[59] profiled 28 circulating amino acids in 140 serum samples from patients with ICC and healthy controls. Six differentially expressed amino acids were applied to develop five machine learning models for ICC diagnosis, achieving accuracies of 73.81%-78.57% and AUCs of 70.83-79.86. Meanwhile, Wu et al[60] applied LASSO, SVM-recursive feature elimination, and RF to two ICC-related human gene expression profiles from the Gene Expression Omnibus database. They identified matrix metalloproteinase-14 as a diagnostic marker and potential therapeutic target due to its role in immune cell infiltration in ICC.

AI IN PROGNOSIS

Currently, adjuvant therapy for ICC remains suboptimal and is often administered to patients with advanced disease, who consequently have a poor prognosis[8-11]. Surgery is the sole curative option for patients with resectable disease. Unfortunately, clinical outcomes following curative resection are disappointing, with a 5-year survival rate ranging from 20%-35%[61]. A meta-analysis of 57 studies (4756 patients) demonstrated that approximately one-third of the patients presented with LNMs (34%), vascular invasion (38%), and/or perineural invasion (PNI, 29%). These adverse pathological factors were associated with postoperative recurrence and shorter OS[61]. Neoadjuvant therapy is considered an effective strategy for downstaging tumors before surgery and may offer potential survival benefits for patients harboring these risk factors[7,62,63]. Collectively, these findings indicate the critical need for more refined preoperative patient management and stratification to identify patients who would benefit maximally from surgery, often necessitating a multimodal treatment approach. Here, the recent advances in AI research for predicting recurrence risk factors (Table 2), ER, and survival outcomes are summarized. Its potential role in facilitating the assessment of treatment response in ICC, encompassing chemotherapy, immunotherapy, and surgery, is also explored.

Table 2 Applications of artificial intelligence in predicting intrahepatic cholangiocarcinoma recurrence risk factors.

Ref.	Sample size	Data source	Algorithms	Aim	Validation/training set results
Xu et al[65]	106 cases	T1-weighted contrast-enhanced MRI	mRMR, SVM	LNM prediction	AUC (0.87), sensitivity (89.47%), specificity (69.57%), accuracy (78.57%)
Xu et al[67]	116 cases	CECT	mRMR, LASSO	TLSs status prediction	AUC (0.85)
Mi et al[68]	271 cases	CEA, CA19-9, number, differentiation, and primary site of tumor	SVM	Prediction of LN status for non-dissected patients	AUC (0.754)
Gao et al[72]	519 cases	DCE-MRI	CNN	MVI prediction	AUC (0.895), sensitivity (73.9%), specificity (89.6%), accuracy (85.6%)
Ma et al[74]	160 cases	T1-weighted MRI, T2-weighted MRI, DWI	LASSO, LR, SVM	MVI prediction	AUC (0.867), sensitivity (64.3%), specificity (80%)
Jiang et al[52]	127 cases	18F-FDG PET/CT	SFFS, RF	MVI prediction	AUC (0.90), sensitivity (75%), specificity (80%), accuracy (77%)
Fiz et al[75]	74 cases	18F-FDG PET/CT	CART, RF	Prediction of MVI and ICC grading	AUC (0.87) for MVI, AUC (0.78) for grading
Liu et al[79]	243 cases	CECT	XGBoost	PNI prediction	AUC (0.831), sensitivity (76.2%), specificity (87.2%), accuracy (81.5%)
Qian et al[82]	178 cases	Gd-DTPA-enhanced MRI	LASSO, LDA	Ki67 prediction	AUC (0.815), sensitivity (75%), specificity (76.7%), accuracy (71.4%)
Peng et al[76]	128 cases	US	SVM, LASSO, LR, GBDT, bagging	Prediction of MVI, PNI, Ki67, VEGF, CK7, and differentiation	Sensitivity (75%), specificity (72.2%), accuracy (73.5%) for ICC differentiation

MRI: Magnetic resonance imaging; CECT: Contrast-enhanced computed tomography; CA19-9: Carbohydrate antigen 19-9; DCE-MRI: Dynamic-contrast-enhanced magnetic resonance imaging; DWI: Diffusion-weighted imaging; 18F-FDG PET/CT: 18F-fluorodeoxyglucose positron emission tomography/computed tomography; Gd-DTPA: Gadolinium-diethylenetriamine pentaacetic acid; US: Ultrasound; mRMR: Maximum relevance and minimum redundancy; SVM: Support vector machine; CNN: Convolutional neural network; LASSO: Least absolute shrinkage and selection operator; LR: Logistic regression; SFFS: Sequential forward floating selection; RF: Random forest; XGBoost: Extreme gradient boosting; LDA: Linear discriminant analysis; GBDT: Gradient boosted decision tree; LNM: Lymph node metastasis; TLSs: Tertiary lymphoid structures; LN: Lymph node; MVI: Microvascular invasion; PNI: Perineural invasion; VEGF: Vascular endothelial growth factor; CK7: Cytokeratin 7; AUC: Area under the curve; CART: Classification and regression tree; CEA: Carcinoma embryonic antigen.

Prediction of recurrence risk factors

LNM: LNM is one of the most critical factors influencing postoperative recurrence and OS in ICC[3,7,12]. Patients with LNM typically exhibit an extremely poor prognosis, and its presence is considered a strong relative contraindication to surgery[3]. A Surveillance, Epidemiology, and End Results database study of 169 ICC patients with positive LNs found a median OS of only 19 months after resection, compared to 20 months with systemic chemotherapy alone[64]. Therefore, accurate preoperative assessment of LN status is crucial for treatment selection. However, current methods for preoperative LN status prediction are often unreliable and suboptimal[18]. Surgical dissection remains the standard strategy for definitive determination of LN status; however, it is highly dependent on the skill and experience of the surgeon. Precise evaluation of LN status non-invasively before surgery is paramount to ensure that patients receive the most appropriate therapy.

Xu et al[65] extracted features from the T1-weighted contrast-enhanced MRI scans of patients with ICC to build an SVM model. They further developed a preoperative nomogram for predicting LN status based on the SVM score, CA 19-9 Level, and MRI reports, achieving an AUC of 0.87 in the validation cohort. Tertiary lymphoid structures (TLSs) are ectopic lymphoid formations that develop in non-lymphoid tissues under chronic pathological conditions, such as autoimmune diseases, chronic infections, and cancer. Recently, TLSs have garnered significant attention due to their potential prognostic value and implications in guiding immunotherapy[66]. Xu et al[67] used maximum relevance and minimum redundancy and LASSO to extract radiomic features predictive of TLS status in patients with ICC. The combined model integrating both clinical and radiomic features demonstrated superior performance (AUC = 0.85) compared with standalone clinical or radiomic models.

The need for routine lymphadenectomy for ICC remains controversial[7]. To evaluate LNM status in patients who did not undergo intraoperative LN dissection or had incomplete dissection, a retrospective study involving 271 patients developed an SVM model that effectively predicted LNM (AUC = 0.754). Furthermore, survival analysis revealed significantly worse OS (P < 0.001) and disease-free survival (P < 0.001) in patients with pathologically confirmed or predicted LNM positivity than those in the predicted LNM-negative group. The research team developed a user-friendly online calculator for practical applications based on this model[68].

Microvascular invasion: Microvascular invasion (MVI) is highly correlated with postoperative recurrence, similar to LNM[3], and is negatively associated with both OS and disease-free survival[69]. Major vascular invasion in primary liver cancer is often visualized preoperatively using different medical imaging modalities. However, detecting MVI is exceptionally challenging because it is a histological finding that is typically identified only in resected pathological specimens[70]. Although some studies have suggested that certain laboratory blood tests, such as alanine aminotransferase, alpha-fetoprotein, and CA 19-9, and radiological features, including tumor size, tumor morphology, and intrahepatic bile duct dilation, possess some predictive capability for MVI[69,71], aspects of these findings remain controversial. Consequently, the ability to predict MVI accurately based on preoperative data can significantly aid clinicians in making optimal treatment decisions.

Radiomics and deep learning studies using different MRI sequences have proven to be powerful tools for predicting MVI in ICC. Among these, Gao et al[72] achieved the highest validation AUC (0.895) using a multiparametric fusion deep learning model based on DCE-MRI. Furthermore, they compared the performance of deep learning models built on two distinct CNN architectures: A multiparametric fusion CNN and a late fusion CNN. Their results indicated that the multiparametric fusion CNN model yielded a superior AUC and accuracy in both the validation and test sets compared to the late fusion CNN model, although the differences were not statistically significant (P = 0.438 and P = 0.504, respectively). Pathologically, MVI predominantly occurs in the peritumoral region, whereas many earlier studies focused solely on the intratumoral features[71,73]. Ma et al[74] investigated how different volumes of interest, defined by radially expanding 8, 10, and 12 mm from the tumor margin, affect MVI prediction. They concluded that radiomic features from the tumor plus a 10-mm peritumoral zone (volume of interest 10 mm) yielded the highest predictive performance (AUC of 0.867). Their results suggest that extracting more radiomic features from the peritumoral region may indicate MVI, implying that wider surgical resection could improve oncological outcomes. Finally, two PET/CT-based radiomic studies have also demonstrated promising predictive capabilities for MVI, reporting AUCs of 0.90 and 0.87, respectively[52,75].

Other recurrence risk factors: In addition to LNM and MVI, PNI, Ki67 protein expression, tumor differentiation, and tumor grade are also significant factors contributing to tumor recurrence and poor prognosis in ICC[61,76]. PNI is characterized by tumor cell infiltration along nerves and/or into the epineurium, perineurium, and endoneurium of the nerve sheath, involving at least one-third of the nerve circumference[77]. As a potential pathway for tumor spread, PNI is recognized as an independent risk factor for the recurrence of ICC and poor long-term survival. Patients with ICC and PNI may benefit from wide resection margins[78]. This underscores the importance of accurate preoperative assessment of PNI to facilitate comprehensive surgical planning by surgeons. Liu et al[79] extracted seven radiomic features from preoperative CECT scans of 243 patients with positive PNI. They then used extreme gradient boosting to build a model integrating these radiomic features with three key clinical-radiological features (platelet-to-lymphocyte ratio, tumor location, and arterial rim enhancement). SHapley Additive exPlanations was used to visualize the prediction process for the clinical applications. In the external validation cohort, this combined model achieved an AUC of 0.831 and accuracy of 81.5%.

Ki67 is a nuclear proliferation-associated antigen whose expression level is directly correlated with tumor aggressiveness[80]. Furthermore, Ki67 is an attractive therapeutic target for ICC treatment. For example, dinaciclib, a small-molecule multi-cyclin-dependent kinases inhibitor targeting cyclin-dependent kinases 2/5/9, can slow ICC progression by inhibiting Ki67 expression[81]. Therefore, accurately predicting Ki67 expression levels in patients with ICC is crucial for assessing treatment efficacy and outcomes. However, determining Ki67 status using conventional imaging and laboratory tests is difficult. Qian et al[82] employed linear discriminant analysis to identify the optimal feature combination for distinguishing between different Ki67 expression levels, including hepatitis B virus status, arterial rim enhancement, enhancement pattern, and contrast-enhanced MRI radiomic features. They constructed a highly efficient Ki67 prediction model (AUC = 0.815).

Tumor differentiation refers to the degree of resemblance between tumor cells and their normal counterparts, as reported by histopathology. Tumors whose cells closely resemble normal cells in morphology or tissue architecture are considered well-differentiated. Conversely, a greater deviation from normal cells indicates poor differentiation, which is often associated with a worse prognosis[61]. Peng et al[76] extracted US radiomic features and used a decision tree to predict the degree of tumor differentiation in ICC, achieving an accuracy of 73.5%, sensitivity of 75%, and specificity of 72.2%. They also evaluated five other prognostic features, MVI, PNI, Ki67, vascular endothelial growth factor, and cytokeratin 7, using SVM, LASSO, logistic regression, gradient boosting decision tree, and bagging algorithms. Building upon cellular differentiation, tumor grade is another pathological concept derived by considering factors such as cellular arrangement, mitotic count, and the extent of local invasion. The tumor grade similarly reflects the malignancy level and prognosis of the tumor. This retrospective study has provided practical radiomic-based tools for predicting tumor grade in ICC[75].

ER prediction

The biological behavior of ICC is extremely aggressive and metastatic, leading to a propensity for local and/or distant recurrence after surgery, which hinders its cure. In addition to the previously mentioned risk factors, recurrence after ICC resection may be associated with tumor size, multiple tumors, tumor burden score, liver cirrhosis, and serum tumor marker levels. Previous studies have developed predictive models for recurrence based on these factors[3,83,84]. However, patients with similar clinical characteristics may exhibit different recurrence patterns, raising questions about the accuracy and generalizability of the predictive models in prior studies[83]. Based on the interval from curative resection to recurrence, recurrence occurring within less than 12 months is typically defined as ER, whereas recurrence beyond 24 months is considered late recurrence[85]. ER is associated with a worse OS. Zhang et al[17] found that, compared to patients with late recurrence (median survival, 18 months), the median survival of ER patients was only 10 months (P = 0.029). Given the significant prognostic implications of ER in poor ICC outcomes and the potential benefits of adjuvant or neoadjuvant therapy for patients with ER[62,86], researchers are motivated to utilize novel technologies to develop more effective models for predicting ER.

Using multivariate logistic regression, Bo et al[87] identified male sex, MVI, tumor-nodes-metastasis stage, and CA 19-9 Levels as independent risk factors for ER. However, the corresponding clinical model demonstrated poor predictive performance (AUC = 0.685). Subsequently, they extracted the 10 most significant features from 57 CECT radiomic features to develop seven machine learning models. These models achieved a mean AUC higher than the clinical model (0.87 ± 0.02), with RF, neural network, and SVM showing the best performance (AUC = 0.89). As previously noted, peritumoral regions of varying widths influence the prediction of MVI in ICC. This may be due to variations in tumor-related information across distinct peritumoral zones. Moreover, Xu et al[88] investigated the impact of different peritumoral region widths (3 mm or 5 mm) using MRI radiomics to predict recurrence. They concluded that for predicting ER, a combined clinical-radiomics model incorporating the intratumoral and the 5 mm peritumoral regions yielded the best performance (AUC = 0.852). Furthermore, they found that radiomic models based on the same regions demonstrated good predictive ability for late recurrence (AUC = 0.735). This highlights the significance of the peritumoral region in radiomics research.

Very ER was first proposed by Zhang et al[17] in 2018. They observed that approximately one-quarter of patients with ICC experienced recurrence within 6 months after the initial resection, indicating a worse prognosis compared to ER. To predict very ER, Chen et al[85] used six different machine learning algorithms to build radiomic-clinical models, achieving a robust predictive performance in an external validation cohort (mean AUC = 0.929). Additionally, based on CECT radiomic features, they classified ICC into two novel subtypes. Subtype 2 exhibits a higher proportion of very ER than subtype 1 (47.62% vs 25.53%) and has a significantly shorter survival time.

Survival prediction

Survival time is one of the most crucial prognostic indicators in patients with malignant tumors. Stratifying patients with ICC according to survival risk is the key to determining the appropriate treatment, particularly for those with advanced ICC. Clinicians typically use the American Joint Committee on Cancer and Union for International Cancer Control staging system to assess ICC aggressiveness. In the latest 8th edition, ICC is staged based on clinicopathological factors such as tumor size, vascular invasion, periductal invasion, and LNM[5,17]. However, several recent studies have indicated that the new tumor-nodes-metastasis staging system does not significantly improve survival prediction[89,90]. Researchers are relying on AI to address this arduous task.

Park et al[91] used DCE-CT radiomic features along with five independent CT features (infiltrative contour, multiple lesions, perivascular infiltration, extrahepatic organ invasion, and suspicious metastatic LNs) to construct a model predicting postoperative recurrence-free survival for MF-ICC. In the test cohort, this model demonstrated a performance similar to postoperatively available prognostic systems, including the 8th edition American Joint Committee on Cancer staging, in predicting both recurrence-free survival and OS. Classification trees, a type of decision tree algorithm designed for classification problems such as assigning samples to different categories or labels, were used by Tsilimigras et al[92]. Using tumor size, CA 19-9 Levels, and neutrophil-to-lymphocyte ratio, patients with ICC were stratified into three distinct prognostic subgroups. The median OS for these subgroups post-resection was 60.4 months, 27.2 months, and 13.3 months, respectively (P < 0.001). As previously discussed, Xu et al[67] and Fiz et al[75] effectively predicted the factors influencing recurrence, namely, LNM and MVI, using radiomic approaches. Furthermore, these studies demonstrated good predictive capabilities for recurrence-free survival and OS.

Histopathological tumor specimens contain more detailed information regarding differentiation and staging than preoperative clinical and radiological data. Yang et al[93] found that combining an MRI radiomics model with a clinicopathological-radiological model incorporating pathological information significantly improved the prognostic accuracy (yielding a net reclassification improvement of 32.5%-34.3%). Xie et al[90] developed a deep learning model based on a deep CNN to explore the predictive value of lymphocyte distribution density and tissue composition in WSIs for ICC survival. The model achieved an AUC of 0.6818 for OS analysis and an AUC of 0.6771 for disease-free survival analysis in an independent test set. Extremely aggressive ICC can lead to postoperative ER or even very ER, sometimes resulting in survival times of less than one year[17]. Identifying such patients preoperatively is critical to avoid potentially futile surgery and seek more suitable treatment options. Müller et al[94] used an artificial neural network to predict 1-year survival after ICC resection, achieving an AUC of 0.80 in the validation cohort.

Treatment response

Owing to its powerful computational capabilities, AI can predict treatment responses to chemotherapy, immunotherapy, and surgery in patients with ICC. For patients with locally advanced disease confined to the liver, intrahepatic chemotherapy via hepatic artery infusion (HAI) delivers agents directly to the liver lesions while minimizing systemic toxicity. Floxuridine is one of the most effective drugs currently used for HAI, achieving higher intratumoral drug levels than systemic administration[95]. Patel et al[96] used unsupervised machine learning to investigate the predictive value of pan-cytokeratin and CD45 expression levels in circulating tumor-immune hybrid cells for response to floxuridine HAI therapy. They discovered that, compared with responsive ICC cells, non-responsive ICC cells exhibited elevated pan-cytokeratin expression and reduced CD45 expression.

Recently, members of the immune checkpoint pathway, such as programmed cell death protein 1 (PD-1) and programmed cell death ligand 1 (PD-L1), have garnered increasing attention. Based on current clinical trials, agents that block the PD-1/PD-L1 pathway have shown significant potential in improving survival in patients with ICC[14]. However, the key to the successful development of immune checkpoint blockade therapies lies in selecting the patient subgroups most likely to benefit, thereby avoiding ineffective treatments and potential side effects associated with autoimmune reactions triggered by blocking PD-1/PD-L1. Ji et al[97] identified three radiotranscriptomic signatures for predicting the response to PD-1/PD-L1-related immunochemotherapy using an RF model, achieving an AUC of 0.84 in the validation cohort. Furthermore, Zhang et al[98] developed an MRI radiomics model to non-invasively predict PD-1 and PD-L1 expression in ICC patients (AUC = 0.84 and 0.89, respectively). They also found that patients who were positive for PD-1 or PD-L1 expression had poorer outcomes than those who were negative. These two studies can assist clinicians in selecting patients with ICC suitable for PD-1/PD-L1 blockade and inform clinical decision-making.

The textbook outcome concept overcomes the limitations of previous single-metric assessment systems by using composite indicators to provide a more comprehensive evaluation of surgical quality. This approach has gained widespread acceptance among surgeons in clinical practice[99]. Huang et al[100] used logistic regression to identify the preoperative variables that influence textbook outcome, including Child-Pugh grade, Eastern Cooperative Oncology Group performance status, hepatitis B infection, and tumor size. They then used the extreme gradient boosting algorithm to build a model capable of preoperatively predicting the textbook outcome (AUC = 0.8346). Post-hepatectomy liver failure and bile leakage are two common surgical complications. Two separate machine learning-based studies have developed predictive models for each complication[101,102].

CHALLENGES AND PROSPECTS

Despite its significant potential for ICC diagnosis and prognosis, AI faces numerous challenges. Data form the foundation of all AI algorithms and models. The volume of data directly affects model effectiveness and credibility, particularly for deep learning models that require large training datasets[22]. As a relatively uncommon malignancy, most prior AI studies related to ICC have been based on small samples, often involving fewer than 500 patients[40,65,75]. Limited data can lead to overfitting and poor performance in external test cohorts. Multi-center data sharing represents an effective solution to the issue of data volume; however, implementing it in real-world settings is challenging. First, there is a deep-rooted reluctance among hospitals to share data, as they view the patient information that they manage as a valuable proprietary asset. Second, the equipment parameters used by different hospitals and the skill levels of their personnel are difficult to standardize. This inherent heterogeneity could introduce bias into the results. Addressing these challenges will require coordinated efforts from national healthcare authorities. Establishing multi-center, large-scale databases with standardized data collection specifications is essential to ensure the generalizability and reliability of intelligent models.

As AI rapidly develops, data privacy concerns continue to grow. Various AI-based applications require continuous patient data input to achieve continuous optimization, often without patients’ awareness, posing privacy risks. For example, in 2017, the United Kingdom Information Commissioner’s Office ruled that Google’s AI subsidiary DeepMind had illegally processed medical records of 1.6 million patients from the United Kingdom National Health Service. This followed a 2015 data-sharing agreement between DeepMind and the Royal Free London National Health Service Foundation Trust, which led to the transfer of sensitive patient data, including human immunodeficiency virus status, drug overdoses, abortion records, and mental health histories, to develop streams, a mobile application for detecting acute kidney injury[103]. To prevent similar privacy breaches in the current digital era, AI professionals must enhance oversight of health data collection and strengthen legal protections for data security.

Radiomics analysis efficiently extracts latent features from tumor images through a complex workflow, enabling rapid development of clinical diagnostic and predictive models. However, this workflow is susceptible to bias and variation owing to the heterogeneity across different centers. Despite the explosive growth in the radiomics literature, some studies frequently fail to adequately account for sources of variation and report results validated only on isolated datasets without external testing. The resulting concerns regarding study quality and reproducibility have slowed the pace of radiomics innovation. Recognizing the need to assess the scientific merits and clinical utility of radiomics research, Lambin et al[104] proposed the radiomics quality score in 2017. This score comprises 16 key items grouped into six domains: Image protocol, feature extraction, data analysis and statistics, model validation, clinical utility, and open science. Each item is assigned a weighted score based on its importance, with a maximum possible score of 36 points. A study evaluating the quality of the radiomics literature found that the median radiomics quality score percentage across 1574 articles was only 31%, indicating significant methodological and scientific shortcomings in most studies. Nevertheless, the quality of radiomics research has improved from 2014 to 2023, with a more pronounced trend towards improvement, particularly after the introduction of the radiomics quality score in 2017[105]. Explainability remains a frequent challenge in deep learning models. Legally, this raises the question of who bears responsibility if a prediction is wrong, because the algorithm cannot justify it[106]. More importantly, the “black box” nature of the model can undermine patient trust. To address this issue, techniques such as class activation mapping have emerged, enabling researchers to examine classified images and understand the parts that contribute most significantly to a model’s final output[107].

Current research on AI systems for diagnosing and prognosing ICC predominantly remains at the development stage, with validation often limited to studies involving small sample sizes. While some investigations have introduced web-based calculators based on proprietary models[68,102], there is a lack of experimental validation regarding the predictive accuracy of these tools. Addressing real-world clinical implementation challenges, Wei et al[38] successfully integrated their LiLNet system into clinical workflows, spanning outpatient, emergency, and inpatient settings, of two hospitals. Upon evaluation, LiLNet demonstrated outstanding performance, achieving AUCs of 0.950 and 0.976 for ICC diagnosis at the respective hospitals. This representative study exemplifies robust validation of AI system feasibility in clinical environments; however, empirical studies remain scarce, with most having retrospective designs[39,67,75].

To establish greater reliability, future efforts should include prospective studies and randomized controlled trials that offer higher levels of evidence. Additionally, the issue of accountability related to incorrect outputs by intelligent systems remains unresolved. Assigning responsibility to physicians is arguably unjust if they were not involved in the system’s development, yet holding developers liable might restrict further innovation in AI algorithms. Another barrier to clinical adoption arises from healthcare professionals’ concerns about job displacement, which can foster distrust and reluctance toward AI-based interventions. Addressing this resistance requires open dialogue regarding the evolving relationship between AI technologies and healthcare professionals.

Finally, large language models, such as ChatGPT and DeepSeek, are rapidly influencing sectors including healthcare. These applications possess the capability to integrate and analyze user-provided medical data to generate comprehensive treatment recommendations. Their primary advantage lies in offering cost-free access to personalized care comparable to that provided by physicians. Nonetheless, ensuring the reliability of these recommendations remains a critical challenge that requires further comprehensive investigations.

CONCLUSION

ICC is a highly malignant primary liver cancer with a poor prognosis. Existing medical technologies face challenges in non-invasively differentiating ICC from HCC, predicting LNM, and identifying ER. AI offers advanced computational methods that provide new tools for diagnosing, differentiating, and evaluating the prognosis of ICC. Addressing challenges in this field may lead to improved patient outcomes in the future.