Artificial intelligence in colorectal liver metastases: From detection and resectability to treatment response and recurrence prediction

Abstract

Colorectal cancer ranks among the commonest malignancies globally, and the liver is the main site of metastatic spread. Approximately 50% of patients develop colorectal liver metastases (CRLM), and its prognosis is strongly influenced by CRLM. The only practical route to cure is resection combined with systemic and locoregional therapy, but the decision-making is interlinked and guided only partially by qualitative imaging and standard clinicopathological tests. In this narrative review, we explore where artificial intelligence (AI) stands across the CRLM pathway: Lesion detection and characterization, deep-learning classification of histopathological growth patterns, automated volumetry and surgical planning, intraoperative guidance, prediction of response to systemic and ablative therapy, and estimation of recurrence and survival. AI performance is often equal to or better than expert performance for development cohorts and captures prognostic insight that is missed by qualitative reading. However, retrospective single-center design, small samples, inconsistent external validation, and fragile feature reproducibility hinder translation. Advances will rely on prospective multi-institutional testing, standardized reporting, privacy-preserving collaborative training, consideration of algorithmic bias and regulatory requirements, and integration of imaging, pathology, molecular, and clinical data. Nevertheless, AI has the potential to be an effective adjunct for interdisciplinary CRLM care.

Key Words: Colorectal liver metastases; Artificial intelligence; Radiomics; Deep learning; Machine learning; Precision oncology

Core Tip: Artificial intelligence is now applied across every phase of colorectal liver metastasis care, but its maturity is uneven. Automated detection and liver volumetry already match expert performance, while models predicting response, recurrence, and survival often flounder during external validation. But the question is no longer whether algorithms can learn clinically relevant patterns, but whether they survive beyond the cohort that produced them. Progress will involve prospective multi-institutional validation, reproducible features, and multimodal data integration that enhances, rather than replaces, multidisciplinary clinical judgment.

INTRODUCTION

Colorectal cancer (CRC) is one of the most frequently diagnosed cancers, and it is responsible for approximately one in ten new cancer diagnoses and a comparable proportion of cancer deaths globally. The liver is the predominant site of distant spread: Approximately 50% of patients with CRC develop colorectal liver metastases (CRLM) synchronously at presentation or metachronously during follow-up. Given that hepatic involvement characterizes the natural history of metastatic CRC, early diagnosis and rational management of CRLM are integral to any prospect of long-term survival[1-3].

The only potentially curative approach is resection. Combined with systemic chemotherapy and ablative approaches in a multidisciplinary setting, resection achieves a 5-year overall survival rate of 40%-45% in properly selected patients. Those results are the outcome of an interrelated cascade of decisions: Identifying and characterizing lesions; evaluating technical resectability and the estimated future liver remnant (FLR); selecting among systemic, surgical, and ablative alternatives; and predicting response and recurrence. Outcomes differ markedly with tumor biology, metastatic burden, molecular profile, and institutional practice, but current clinical criteria account for them only partially[3,4].

In conventional practice, we rely on qualitative radiological interpretation and a few clinicopathological variables. Both are susceptible to interobserver variation and have modest predictive value. Most of the biological details hidden in routine cross-sectional images are not observable by the eye and are not used. Radiomics involves extraction of quantitative features at high throughput from standard-of-care images and is combined with machine learning. Therefore, it represents a means of translating these images into reproducible and mineable data to support diagnostic, predictive, and prognostic decisions[4,5].

Against this background, AI, including radiomics, machine learning, and deep learning, has been applied to the entire clinical course of CRLM: Detection on computed tomography (CT) and magnetic resonance imaging (MRI), histopathological characterization, resectability assessment and surgical planning, intraoperative guidance, prediction of response to systemic and locoregional therapy, and prediction of recurrence and survival. This narrative review integrates the evidence across that continuum, evaluates methodological quality and clinical preparedness, and highlights barriers to translating these findings into routine use, namely validation, generalizability, algorithmic bias, regulation, and workflow integration[2]. Figure 1 presents a conceptual framework of AI for the CRLM care pathway.

Figure 1 Conceptual framework for artificial intelligence across the colorectal liver metastasis care pathway. The pathway is shown as five sequential domains: A: Detection and characterization; B: Resectability assessment and surgical planning; C: Intraoperative guidance; D: Treatment-response prediction; E: Recurrence and survival prediction. Input data types (computed tomography, magnetic resonance imaging, whole-slide histopathology, molecular and clinical variables) feed the corresponding artificial intelligence methods (radiomics, machine learning, deep learning), and each domain is annotated with its current stage of clinical maturity. 3D: Three-dimensional; CRLM: Colorectal liver metastases; CT: Computed tomography; FLR: Future liver remnant; ICG: Indocyanine green; MRI: Magnetic resonance imaging; RFS: Recurrence-free survival.

METHODS

This article is a narrative review. No systematic review protocol was registered or followed. We searched PubMed/MEDLINE from database inception until 30^th June 2026 for English-language articles reporting AI in CRLM. The search combined AI-related concepts (“artificial intelligence”, “machine learning”, “deep learning”, “neural network”, and “radiomics”) with the disease term “colorectal liver metastases” and was further extended with task-specific search terms, including detection, segmentation, resectability, treatment response, chemotherapy, ablation, recurrence, and survival.

The primary search yielded 190 unique records. We screened titles and abstracts for relevance to AI applications in CRLM, and records discussing non-hepatic disease, describing no AI methodology, or available only as conference abstracts were excluded. Next, we manually searched the reference lists of retrieved studies and recent reviews, including foundational and methodological sources on radiomics standardization, reporting guidelines, algorithmic bias, federated learning, and multimodal integration that a disease-specific search does not retrieve. Forty-one publications were included in the review: 13 from the direct search and 28 from manual searching and targeted retrieval of contextual or methodological work.

We employed no formal eligibility criteria, quality scoring, or quantitative synthesis because the review is narrative. Selection reflects the authors’ assessment of relevance and clinical applicability and preferentially focuses on recent original investigations and studies with external validation. For this reason, we did not develop a PRISMA flow diagram: Selection was a purposive expert-panel process rather than a replicable screening cascade, and a flow diagram would suggest methodological rigor that the design does not lend itself to. The search yield and provenance of the cited sources are reported above to allow readers to assess the coverage directly.

AUTOMATED DETECTION AND LESION CHARACTERIZATION

CRLM detection via cross-sectional imaging is the task on which everything downstream relies, and it has been the predominant focus of AI studies. A deep-learning lesion-detection algorithm yielded a per-lesion sensitivity of approximately 82% for CT, comparable to the performance of abdominal radiologists and residents, while producing more false positives per patient. That profile suggests an assistive rather than an independent role. A wider synthesis of 17 studies and more than 14000 patients, most of whom had CRC, revealed CT as the most frequently studied modality and found diagnostic accuracies above 90% for multiple deep learning and radiomics models. Models are predominantly applied to the detection of CRLM, but performance may differ with cohort and acquisition protocol[6,7].

A potential line of work involves predicting metastatic disease before radiological discovery. In a proof-of-concept study utilizing CT radiomics with formal verification methods, features extracted from tumor-free or benign-appearing liver parenchyma were used to identify patients who would eventually develop CRLM with high precision, albeit in a very small and heterogeneous sample. Deep-learning image reconstruction technology has been applied separately to preserve the conspicuity of lesions of 5 mm or less on reduced-dose spectral CT, which minimizes radiation exposure during the repeated imaging required for surveillance. Both findings indicate that AI can widen the diagnostic window and reduce the follow-up burden, and both need to be confirmed in much larger prospective cohorts[8,9].

Beyond detection, AI is also applied to lesion characterization and the identification of prognostic tissue characteristics. Histopathological growth patterns (HGPs), especially the distinction between desmoplastic and non-desmoplastic tumor-liver interfaces, are among the best predictors of outcome after resection, but manual characterization is laborious and subject to interobserver variation. A deep-learning method implemented using neural image compression distinguished between desmoplastic and non-desmoplastic HGPs on hematoxylin and eosin-stained whole-slide images [area under the curve (AUC) = 0.93 in development and 0.95 in external validation]. The automated scores recapitulated the survival associations obtained by manual measurement and established AI not as a means of detection alone but as a path to standardized, reproducible characterization[10].

Collectively, these studies suggest that AI can achieve expert performance in CRLM detection and enable reproducible measurement of tissue-level phenotypic properties that are difficult to quantify otherwise. The common deficiencies are equally apparent: Predominantly single-center retrospective designs, small sample sizes, and inconsistent external validation. Whether these advances lead to earlier diagnosis and improved prognosis will need to be investigated prospectively in multi-institutional settings, with resectability assessment as the first test[7,10].

RESECTABILITY ASSESSMENT AND SURGICAL PLANNING

Technical resectability turns largely on whether an adequate FLR can be preserved, which, in turn, depends on reliable volumetry of the liver and its anatomical segments. Manual segmentation is laborious and operator-dependent, hence the interest in automation. A three-dimensional (3D) deep-learning model performed fully automated segmentation of Couinaud segments and the FLR on contrast-enhanced CT, achieving a Dice similarity coefficient of 0.93-0.95 on external validation. The resulting FLR and FLR-percentage estimates did not differ significantly from manual measurements and reproduced the same indications for major hepatectomy. The approach was later extended to automated segmentation of the hepatic and portal veins, enabling blood-free FLR assessment and bringing vascular anatomy, which defines resection planes, within the scope of automated planning[11,12].

AI may also assist surgical decision-making by predicting postoperative risk and oncological outcome and help refine patient selection. In a multicenter cohort, a random-forest model integrating demographic, clinical, laboratory, and genetic variables, including KRAS, BRAF, and mismatch repair status, predicted postoperative complications and survival among patients undergoing simultaneous colorectal and liver resection. The model achieved moderate discrimination, yielded favorable results on decision-curve analysis, and was released as a web-based calculator. Data-driven approaches of this kind have been proposed as alternatives to conventional clinical risk scores after CRLM resection, although their incremental advantage over established scores remains modest and unconfirmed in routine practice[13,14].

Automated volumetry is close to clinical applicability, with accuracy approaching expert manual measurement. Surgical risk prediction is in an earlier stage of development. The limitations mirror those seen in lesion detection: Single-center cohorts, little prospective evaluation, and uncertain generalizability across scanners, populations, and surgical practices. Prospective multi-institutional studies will be needed to show that these tools genuinely change operative decisions and improve outcomes[11,13].

INTRAOPERATIVE GUIDANCE AND ROBOTIC SURGERY

Preoperative planning only helps if it survives contact with the operating field, and the intraoperative setting is where AI in CRLM is least developed but arguably most interesting. Automated segmentation underpins the growing use of 3D reconstructions for operative planning and navigation. Reconstructions can be combined with augmented reality and indocyanine green (ICG) fluorescence so that preoperative anatomical understanding and intraoperative guidance reinforce one another: More faithful reconstruction improves the plan, while better navigation makes the plan reproducible at the table. In minimally invasive liver surgery, this pairing may help translate complex planning into precise parenchymal-sparing resection, although the evidence remains largely descriptive and single-center rather than comparative[15].

Work specific to CRLM has begun to move beyond description. Combining real-time ICG administration with computer vision, investigators classified fluorescence patterns intraoperatively and delineated CRLM from surrounding parenchyma, distinguishing malignant from benign tissue on the basis of dye behavior rather than the impression of the surgeon alone. The practical appeal is obvious: Subcapsular lesions that are invisible or ambiguous on inspection and palpation, a recurring problem after neoadjuvant chemotherapy and in laparoscopic work where tactile feedback is absent, become candidates for objective identification. The series was small and exploratory, and the technique will need prospective evaluation against intraoperative ultrasound before it can be judged[16].

Similar ideas are being tested for anatomical orientation. A color-coded system applied AI to laparoscopic video to label liver segments and vascular structures in real time, and an AI-enhanced navigation model has been used during robotic hepatectomy to flag the inferior vena cava and the roots of the major hepatic veins earlier than they would ordinarily be recognized. Robotic platforms are a natural home for this kind of augmentation: The operative field is already fully digitized, the camera is stable, and overlays can be rendered without adding instrumentation. Whether such systems reduce vascular injury, conversion, or blood loss is unknown; the published experience consists of preliminary reports and technical notes, not comparative trials[17,18].

For the moment, intraoperative AI for CRLM should be considered as a research direction rather than a body of evidence. Existing reports are from proof-of-concept studies, the denominators are small, and none addresses oncological outcome. The requirement is the same for this field as elsewhere: Prospective, multicenter evaluation with outcomes that matter to patients rather than to engineers. However, it is more difficult to satisfy in the operating theater.

PREDICTION OF TREATMENT RESPONSE

Treatment response is conventionally assessed with the Response Evaluation Criteria in Solid Tumors (RECIST), which measures change in lesion diameter and therefore recognizes benefit only once morphological shrinkage has occurred. AI can offer earlier and more granular predictions straight from imaging. A deep-learning radiomics model built on a residual convolutional neural network and applied to contrast-enhanced CT predicted response to first-line chemotherapy with an AUC of 0.82 in validation, outperforming both a conventional handcrafted radiomics classifier and the carcinoembryonic antigen (CEA) level. A delta-radiomics signature based on textural change between baseline and early follow-up CT predicted response to first-line oxaliplatin-based chemotherapy at the level of the individual lesion and correctly reclassified many metastases that RECIST had labelled as responding. These are important results: They show that AI can predict systemic-therapy response earlier and more precisely than size-based criteria[19,20].

AI has also been used to read morphological rather than purely dimensional change. In patients with unresectable liver-only metastatic CRC treated with bevacizumab-based chemotherapy, a residual neural network trained to grade morphological response reclassified a substantial minority of radiologist-designated non-responders as responders. Those reclassified patients had significantly longer overall survival and were more often down staged to curative-intent treatment. By capturing treatment-induced changes in lesion architecture that conventional interpretation misses, such models may identify candidates for conversion to resection who would otherwise be overlooked, and this links response assessment directly to surgical decision-making[21].

Locoregional therapy has attracted similar attention, since predicting the durability of local control is central to patient selection. A machine-learning model using pre-ablation CT radiomic features predicted local tumor progression (LTP) after thermal ablation with a concordance index of 0.79, outperforming a clinical-variable model. When three previously published radiomics models for post-ablation LTP were tested without retraining on independent internal and external cohort data, discrimination fell to near-chance, with concordance indices of roughly 0.47-0.50 against 0.78 in the original report. The authors concluded that LTP cannot yet be predicted reliably from ablation-zone radiomics. The contrast between the two studies is instructive, and it is the single most useful cautionary result in this literature[22,23].

The treatment-response literature thus displays the promise and central weakness of radiomics in CRLM. Models perform well for cohorts whose data they were trained on and may detect response earlier and at finer resolution than RECIST, yet their apparent accuracy frequently fails to survive independent validation, evaluations exposing overfitting, small samples, and heterogeneity in imaging and segmentation. Until prediction models demonstrate generalizability across institutions, their role in guiding systemic and locoregional therapy remains investigational. The same caution applies to recurrence and long-term survival[20,23].

PREDICTION OF RECURRENCE AND SURVIVAL

Most patients with CRLM have recurrence even after curative-intent resection, and accurate prognostication matters for surveillance intensity, adjuvant therapy, and counselling. Traditional tools such as the Fong Clinical Risk Score discriminate poorly, which has encouraged data-driven alternatives. For a multicenter cohort of 572 patients, a machine-learning model incorporating clinical, laboratory, and tumor variables, including nodal stage, CEA, tumor diameter, extrahepatic disease, and KRAS status, predicted overall and recurrence-free survival after hepatectomy with a concordance index of 0.65, against 0.57 for the Fong score. The performance was moderate, but the model demonstrated that machine learning can fold diverse predictors into individualized estimates[24].

Imaging-based models extend prognostication beyond tabular clinical data. A deep-learning network trained on multi-sequence MRI predicted one-year recurrence and recurrence-free survival after CRLM resection, with an AUC of 0.84 and a concordance index of 0.73; high-risk patients carried an approximately four-fold greater hazard of recurrence. For recurrent disease treated with microwave ablation, a CT radiomics nomogram predicted subsequent LTP and separated patients into groups with significantly different progression-free survival. Quantitative imaging appears to carry prognostic information beyond established clinicopathological characteristics[25,26].

An interesting direction uses imaging as a non-invasive window onto tumor biology. Applying an interpretable machine-learning model to preoperative CT, investigators derived a radiomic score estimating tumoral expression of CD73, an ectoenzyme that generates immunosuppressive adenosine and whose high expression is associated with early recurrence. This radiomic CD73 surrogate independently predicted shorter time to recurrence and poorer disease-specific survival after resection over and above the standard clinical risk score. This finding suggests that radiomics may serve both as a non-invasive prognostic marker and an indicator of responsiveness to adenosine-pathway immunotherapies[27].

Histopathology offers another substrate for prognostic modelling. In a prospective study, a transformer-based deep-learning model classifying HGPs on whole-slide images confirmed that desmoplastic tumors were associated with significantly longer overall and progression-free survival than non-desmoplastic tumors and enabled automated pattern-based risk stratification. Model assistance also improved the diagnostic accuracy of junior pathologists and shortened reading time, a workflow benefits distinct from the prognostic signal itself[28].

In summary, prognostic models built on clinical, radiological, molecular-surrogate, and histopathological data generally outperform classical risk scores for their development cohorts, and the convergence of these complementary signals points towards integrated, multimodal prognostication. The same caveats persist: Most models are retrospective, modest in scale, and insufficiently validated across institutions, and few have been tested prospectively or shown to change management[24,28]. Representative studies across detection, resectability assessment, treatment-response prediction, and recurrence prediction are summarized in Table 1, and the corresponding model inputs, performance metrics, and validation status are set out in Table 2.

Table 1 Representative artificial intelligence applications across the colorectal liver metastasis continuum.

Ref.	Model and modality	Key metric	Principal finding or limitation	Maturity
Detection and characterization
Kim et al[6], 2021	Deep learning lesion detection; contrast-enhanced CT	Sensitivity approximately 82% per lesion	Comparable to radiologists but with more false positives; suited to an assistive role	Mature
Höppener et al[10], 2024	Deep learning (neural image compression); H&E whole-slide images	AUC: 0.93/0.95 (dev/external)	Reproducible desmoplastic vs non-desmoplastic growth-pattern classification	Emerging
Resectability and surgical planning
Xie et al[11], 2023	3D deep learning segmentation; contrast-enhanced CT	DSC 0.93-0.95; FLR approximately manual	Automated couinaud-segment and future-liver-remnant volumetry reproducing hepatectomy indications	Mature
Chen et al[13], 2024	Random forest; clinical and genetic variables	AUC approximately 0.70-0.74 (external)	Predicts complications and survival after simultaneous resection; deployed as a web tool	Emerging
Intraoperative guidance
Hardy et al[16], 2023	Computer vision; real-time ICG fluorescence video	Malignant vs benign discrimination	Intraoperative delineation of CRLM from surrounding parenchyma; small exploratory series	Early
Nakano et al[18], 2025	AI-enhanced navigation; robotic hepatectomy video	Technical report; no comparative metric	Earlier intraoperative identification of the IVC and major hepatic vein roots	Early
Treatment-response prediction
Wei et al[19], 2021	Deep learning radiomics (ResNet); contrast-enhanced CT	AUC 0.82 (validation)	Outperformed handcrafted radiomics and CEA for chemotherapy-response prediction	Emerging
Taghavi et al[22], 2021 and van der Reijd et al[23], 2024	CT radiomics; pre-ablation	C-index 0.79 vs approximately 0.47-0.50 (dev vs external)	Strong internal performance not reproduced externally; illustrates the generalizability gap	Early
Recurrence and survival prediction
Tang et al[25], 2024	Multi-sequence MRI deep learning	AUC 0.84; c-index 0.73	Predicts 1-year recurrence and recurrence-free survival after resection	Emerging
Saber et al[27], 2023	Interpretable ML radiomics; contrast-enhanced CT	Independent predictor of TTR and DSS	Noninvasive imaging surrogate of CD73 expression with independent prognostic value	Early
Lam et al[24], 2023	Machine learning clinical model	c-index 0.65 vs Fong 0.57	Outperforms a classical clinical risk score for post-hepatectomy prognostication	Emerging

Maturity categories: Mature, clinically practicable and near expert-level; Emerging, promising but requiring broader validation; Early, proof of concept or failed external validation. 3D: Three-dimensional; AI: Artificial intelligence; AUC: Area under the receiver operating characteristic curve; CEA: Carcinoembryonic antigen; c-index: Concordance index; CRLM: Colorectal liver metastases; CT: Computed tomography; dev: Development cohort; DSC: Dice similarity coefficient; DSS: Disease-specific survival; FLR: Future liver remnant; H&E: Haematoxylin and eosin; ICG: Indocyanine green; IVC: Inferior vena cava; ML: Machine learning; MRI: Magnetic resonance imaging; ResNet: Residual convolutional neural network; TTR: Time to recurrence.

Table 2 Model inputs, performance, and validation status of representative artificial intelligence models in colorectal liver metastases.

Ref.	Task	Input data	Performance	Validation status
Kim et al[6], 2021	Lesion detection	Contrast-enhanced CT images	Sensitivity approximately 82% per lesion	Internal only; retrospective, single-centre
Höppener et al[10], 2024	Growth-pattern classification	H&E whole-slide images	AUC 0.93 (dev); 0.95 (external)	External validation performed
Xie et al[11], 2023	Couinaud segment and FLR segmentation	Contrast-enhanced CT images	DSC 0.93-0.95; FLR approximately manual	External validation performed
Xie et al[12], 2023	Blood-free FLR assessment	Contrast-enhanced CT images; hepatic and portal veins	Automated FLR approximately manual reference	Internal validation; single-centre
Chen et al[13], 2024	Postoperative complications and survival	Demographic, clinical, laboratory and genetic variables (KRAS, BRAF, mismatch repair)	AUC approximately 0.70-0.74	External, multicentre validation
Hardy et al[16], 2023	Intraoperative lesion delineation	Real-time ICG fluorescence video	Malignant vs benign discrimination reported	Exploratory; single-centre; no external cohort
Wei et al[19], 2021	Chemotherapy-response prediction	Contrast-enhanced CT images	AUC = 0.82	Internal validation; single-centre
Giannini et al[20], 2022	Per-lesion chemotherapy response	Baseline and early follow-up CT (delta features)	Reclassified lesions misclassified by RECIST	Internal validation; multicentre cohort
Chiu et al[21], 2024	Morphological response and survival	Contrast-enhanced CT images	Reclassified non-responders had longer overall survival	Internal validation; single-centre
Taghavi et al[22], 2021	Local tumour progression after ablation	Pre-ablation CT radiomic features	c-index 0.79	Internal validation only
van der Reijd et al[23], 2024	Local tumour progression after ablation	Pre-ablation CT radiomic features	c-index approximately 0.47-0.50	Independent internal and external validation; discrimination lost
Lam et al[24], 2023	Overall and recurrence-free survival	Clinical, laboratory and tumour variables	c-index 0.65 (Fong score 057)	Internal validation; multicentre
Tang et al[25], 2024	One-year recurrence and RFS	Multi-sequence MRI	AUC = 0.84; c-index 0.73	Internal validation
Saber et al[27], 2023	CD73 expression, TTR and DSS	Preoperative contrast-enhanced CT	Independent predictor of TTR and DSS	Internal validation; single-centre
Lin et al[28], 2025	Growth-pattern classification and risk stratification	H&E whole-slide images	Improved junior pathologist accuracy; reduced reading time	Prospective, multicentre

AUC: Area under the receiver operating characteristic curve; c-index: Concordance index; CT: Computed tomography; dev: Development cohort; DSC: Dice similarity coefficient; DSS: Disease-specific survival; FLR: Future liver remnant; H&E: Haematoxylin and eosin; ICG: Indocyanine green; MRI: Magnetic resonance imaging; RECIST: Response Evaluation Criteria in Solid Tumors; RFS: Recurrence-free survival; TTR: Time to recurrence.

CLINICAL INTEGRATION: WORKFLOW AND MULTIDISCIPLINARY DECISION-MAKING

Discussion of AI in CRLM usually turns on accuracy, yet accuracy is not the only thing a busy service buys. Much of the near-term value may lie in workload. Manual segmentation of Couinaud segments and the FLR is the clearest example: It takes an experienced operator a substantial part of an hour per case, it is repeated whenever the plan changes, and it varies between operators. Automated volumetry returns measurements that reproduce manual indications for major hepatectomy within seconds, which converts volumetry from a rationed favor into something that can be requested for every candidate and repeated after portal vein embolization at no meaningful cost[11,12]. The pathology literature supplies the only direct measurement of this effect in CRLM: Model assistance both improved the accuracy of junior pathologists in classifying HGPs and reduced their reading time, allowing the same tool to improve quality and reduce effort simultaneously[28].

Framed this way, several applications reviewed here are better understood as triage or standardization than as diagnosis. A detection algorithm that matches radiologists but over-calls false positives is unattractive as an autonomous reader and useful as a second look, particularly for surveillance studies where the reader already knows what to expect and small metastases are easily missed[6]. Deep-learning reconstruction that maintains conspicuity at reduced dose changes what surveillance costs the patient rather than what it detects[9]. These are unglamorous benefits, and they are also the ones most likely to be realized first, because they do not require the algorithm to be trusted with a decision.

The multidisciplinary team (MDT) meeting is where any of this would actually bite. CRLM management is decided in that room, and the decisions are exactly the ones the models address: Is this resectable? Will this respond? Is this patient likely to have early recurrence? A systematic review of AI-enhanced oncology MDTs reported gains in preparation efficiency and the consistency of recommendations and reduced time to decision while noting that the evidence base is young and dominated by single-center implementations[29]. Large language models have been evaluated by tumor boards with much the same conclusion: Concordance with board recommendations is reasonable for standard cases and degrades for complex or atypical ones, which is precisely the population an MDT exists to discuss[30]. A model that agrees with the team when the answer is easy adds little; the value would lie in flagging the difficult case, and that has not yet been demonstrated.

Two practical obstacles are notable. The first is trust calibration. Tools such as the web-based risk calculator derived from the simultaneous-resection cohort are easy to consult in a meeting, which is both their advantage and their hazard: A number presented on a screen carries an authority its moderate discrimination does not warrant, and there is no mechanism for a typical MDT for conveying that a concordance index of 0.65 means the model is often wrong about individuals[13]. Saliency maps and similar post-hoc explanations are frequently offered as the remedy, but they are unreliable, can appear plausible while the underlying model is wrong, and should not be treated as a substitute for validation[31]. The second is accountability. An MDT recommendation is attributable to the team; an algorithmic recommendation is attributable to nobody in the room. Until that is resolved by governance rather than by software, an AI output belongs in MDT discussions as one more imperfect input, in the same category as a borderline CEA or an equivocal biopsy.

Shared decision-making with patients is the least examined aspect of all. The decisions at stake, whether to accept a major hepatectomy, whether to continue chemotherapy hoping for conversion, how intensively to be followed, depend on how a patient weighs the chance of cure against morbidity, and individualized estimates could in principle support that conversation better than a five-category risk score. In principle, communicating a model-derived probability honestly requires conveying its uncertainty and its provenance, including the awkward fact that a model derived elsewhere may not apply to the patient in front of you. Evidence on how patients with CRLM interpret such estimates is lacking, and there is no reason to assume that the transfer from clinician to patient is straightforward.

CHALLENGES, LIMITATIONS, AND FUTURE DIRECTIONS

The applications reviewed here share constraints that impede translation. Most studies are retrospective and single-center. They also enroll relatively small cohorts, and report performance only for internal or development datasets. As the ablation literature shows, apparent accuracy often does not hold up under independent external validation. Fixing this requires adequately powered prospective, multi-institutional studies alongside standardized reporting and evaluation. Dedicated frameworks, including the artificial intelligence extensions to the Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis and the Prediction model Risk of Bias Assessment Tool, were developed to improve transparency, support critical appraisal, and reduce avoidable research waste in AI prediction-model studies[32].

A closely related problem is the reproducibility of the features on which radiomic models rest. Quantitative imaging features are sensitive to scanner hardware, acquisition and reconstruction parameters, segmentation method, and software implementation, and nominally identical features can vary substantially across center. The Image Biomarker Standardization Initiative addressed this by defining and validating a set of standardized radiomic features for calibration and verification across software platforms. Comparable standardization is needed across the wider colorectal radiomics pipeline before models can be expected to transfer reliably[33,34].

Algorithmic bias has received little attention in the CRLM literature and deserves more. Models learn from the populations on which they are trained, and the cohorts behind the studies reviewed here are overwhelmingly single-center and drawn from high-income European, North American, and East Asian settings. A model with training data from a few patients with steatotic or chemotherapy-injured livers or a few patients from a given ethnic group may perform worse for these patient populations without anything in its reported accuracy revealing the shortfall. The mechanism is not hypothetical: A widely deployed population health algorithm was shown to systematically under-refer Black patients because it used healthcare expenditure as a proxy for illness, and expenditure encoded unequal access rather than need[35]. Radiomic pipelines are vulnerable to the same class of error whenever a convenient label stands in for the thing of real interest, as when radiological response substitutes for benefit. Reporting cohort composition and testing performance within subgroups should be routine, and at present it is not.

The opacity of deep learning compounds this. A radiomic score predicting CD73 expression or a network grading morphological response offers no account of why a given patient was classified as it was, which matters when the output contradicts clinical judgment. Post-hoc explainability methods are widely promoted as the answer, but current approaches are unreliable and can lend spurious credibility to a flawed model; rigorous external validation, not a heat map, is what establishes trustworthiness[31]. Regulation is a further hurdle that CRLM tools have barely begun to encounter. An analysis of AI devices cleared by the United States Food and Drug Administration found that most were evaluated retrospectively, many at a single site, and few were assessed prospectively or reported performance across patient subgroups[36]. Almost none of the models discussed in this review would currently satisfy a stricter standard, and the gap between publication and deployment is largely this evidentiary one rather than a technical one.

Data scale and privacy are also limiting. Multi-institutional collaboration is essential because single institutions rarely hold datasets large or diverse enough to train robust models, yet centralized pooling raises privacy and governance problems. Federated learning offers a route around this by distributing training to the data owners and aggregating only model updates, and patient-level data never leave the institution. In a study spanning 10 institutions, federated models reached approximately 99% of the performance of a model trained on centrally pooled data[37]. The approach is well suited to CRLM specifically, where the obstacle to progress is not algorithmic novelty but the difficulty of assembling enough imaging, pathology, and outcome data across centers to test whether a model generalizes and where the barrier to sharing is usually legal rather than technical.

Federated learning is not a panacea, and it is worth being explicit about what it does and does not solve. It addresses data governance. It does not by itself address bias, since a federation of similar institutions merely aggregates a shared blind spot. It also does not remove the need for harmonized acquisition and segmentation, because heterogeneous inputs degrade a federated model just as they degrade a centralized one. It introduces its own difficulties: Uneven data quality between sites, communication overhead, contributions weighted by cohort size in ways that can let one large center dominate, and the residual privacy risk that model updates themselves can leak information. The infrastructure and governance requirements are substantial, and the participation incentives for smaller centers are not obvious[38]. For CRLM, the practical target is modest and achievable: A federated consortium able to test whether an ablation or recurrence model derived at one center holds at twenty others, which is precisely the question the current literature cannot answer.

The most substantial gains are likely to come from integrating the data streams this review has discussed separately. Current CRLM models are almost entirely unimodal and use imaging, histopathology, molecular data, or clinical variables in isolation, whereas the biology and management of the disease are inherently multimodal[39]. Concrete examples now exist elsewhere in oncology. A pan-cancer analysis of 14 cancer types trained a model to combine whole-slide histology with molecular profiles, and the multimodal model outperformed both unimodal versions for prognostic stratification while allowing the contribution of each data type to be interrogated[40]. Broader surveys of multimodal AI in oncology describe the same pattern across imaging, pathology, genomics, and electronic health records, and identify the practical obstacles: Missing modalities in real cohorts, the need for spatial and temporal alignment between data types, and sample sizes that must be larger than any single modality requires[41]. Applied to CRLM, a model that combined preoperative CT, resection-specimen HGPs, KRAS and mismatch repair status, and routine laboratory values would mirror how an MDT actually reasons; the reason it does not yet exist is that no single center holds enough patients with all four, which is exactly the gap federated collaboration is meant to close.

Overall, AI has shown measurable potential across the CRLM pathway, from detection and characterization through resectability assessment, intraoperative guidance, treatment-response prediction, and prognostication, but it remains largely investigational. Routine implementation will require prospective multi-institutional validation, standardized features and transparent reporting, explicit attention to bias and regulatory expectations, privacy-preserving collaborative training, explainable and workflow-compatible tools, and ultimately multimodal integration. If those conditions are met, AI may develop from a promising research tool into a dependable adjunct within the multidisciplinary care of patients with CRLM[32,39].

NOVELTY OF THE STUDY

Existing reviews of artificial intelligence for CRLM tend to treat the field one task at a time, most often detection or segmentation, and tend to report performance approvingly. This review differs in three respects. First, it follows the disease along its actual clinical pathway, from detection and characterization through resectability assessment, intraoperative guidance, treatment-response prediction, and prognostication, and grades each application by clinical maturity rather than by reported accuracy, so that tools ready for use are separated from those that are not.

Second, it treats failure as evidence. The finding that published ablation radiomics models collapse to near-chance discrimination on independent testing is given the same weight as the headline results, because the gap between development and external performance is the central obstacle in this field and is routinely underplayed. Third, it extends the discussion into territory the CRLM literature has largely left alone: The workload and workflow consequences of these tools, their behavior in multidisciplinary and shared decision-making, algorithmic bias in cohorts that are narrow in ways their accuracy figures do not reveal, regulatory expectations, and what federated and multimodal approaches can and cannot fix.

The intended contribution is a candid appraisal of readiness rather than an inventory of achievements: A map of where artificial intelligence in this disease is usable now, where it is promising, where it has already failed, and what evidence would be required to move each application from one category to the next.

CONCLUSION

Artificial intelligence is reshaping the detection, characterization, and management of CRLM, but its maturity varies considerably across the care continuum. Algorithm-based lesion detection and liver volumetry have progressed the farthest, approaching expert-level performance and practical applicability, whereas models predicting treatment response, recurrence, and survival remain earlier in development and more vulnerable to the generalizability problems that run through this literature. Intraoperative and robotic applications are still new and rest at present on proof-of-concept reports rather than comparative outcome data. The key question is no longer whether AI can learn clinically relevant patterns, which it clearly can, but whether those patterns remain valid beyond the cohorts whose data were used for training. Bridging that gap will depend less on developing increasingly complex algorithms than on prospective multi-institutional validation, consistent and reproducible inputs, transparent reporting, honest handling of bias and uncertainty, and seamless integration into multidisciplinary workflows. With progress in these areas, and ultimately through the incorporation of multimodal data, AI is positioned to become an effective decision-support tool that complements rather than replaces clinical judgment and supports more informed management of patients with CRLM.