Artificial Intelligence and Radiomics for Predicting Chemotherapy Response in Colorectal Liver Metastases: A Systematic Review and Meta-Analysis

Colorectal cancer (CRC) remains one of the most prevalent malignancies worldwide and a leading cause of cancer-related mortality. ¹ A significant proportion of patients with CRC, which is approximately 25-50%, will develop liver metastases either at the time of diagnosis or during the course of disease progression. ^1,2 These colorectal liver metastases (CRLM) represent a critical therapeutic challenge, as their presence significantly impacts prognosis, with 5-year survival rates remaining below 15% without intervention. ³ The mainstay of CRLM management includes systemic chemotherapy, targeted therapies, and surgical resection, with the latter being the only potentially curative option. ² However, not all patients respond equally to systemic treatments, and early identification of responders versus nonresponders is essential for tailoring therapeutic plans, minimizing toxicity, and maximizing survival benefit. ⁴

Traditional methods of evaluating treatment response rely heavily on morphological criteria such as the Response Evaluation Criteria in Solid Tumors (RECIST) version 1.1. ⁵ While RECIST provides standardized guidelines for assessing tumor shrinkage, it is inherently limited by its dependence on linear size measurements and its inability to capture subtle biological or functional changes within tumors. ^6,7 Consequently, there has been growing interest in more sophisticated tools that can quantify intratumoral heterogeneity and predict treatment outcomes beyond size-based assessments.

Radiomics refers to the high-throughput extraction of quantitative features from medical imaging data. These features, encompassing tumor shape, intensity, texture, and wavelet transformations, reflect underlying pathophysiological processes and can provide insights into tumor behavior, aggressiveness, and response to therapy. ^8-11 When paired with artificial intelligence (AI) techniques such as machine learning (ML) and deep learning (DL), radiomics can be transformed into predictive models that recognize complex patterns and make data-driven predictions. ML algorithms can be trained on large, labeled datasets to distinguish between responders and nonresponders based on imaging-derived features, while DL approaches such as convolutional neural networks can autonomously learn hierarchical representations directly from raw imaging data. ¹²

The application of AI and radiomics in CRLM has gained traction in recent years, driven by advances in computational power, access to large imaging datasets, and improvements in imaging standardization. Numerous studies have explored the use of radiomics-based signatures derived from CT, MRI, or PET/CT to predict histopathologic response or RECIST-based radiologic response in patients undergoing chemotherapy. These studies vary widely in terms of patient population, treatment regimens, feature extraction methods, and validation approaches, necessitating a systematic synthesis to determine overall diagnostic performance. ^13-33

This meta-analysis aims to address this gap by pooling data from studies using radiomics-based AI models to predict chemotherapy response in CRLM. In addition to evaluating overall accuracy, we assess model performance in key subgroups, particularly those receiving chemotherapy alone versus those receiving combination therapy with biologics such as bevacizumab or cetuximab.

There are 21 studies included in this systematic review and meta-analysis. These were published between 2020 and 2025. Overall, there are 2769 patients included in this study. ^13-33 All studies used radiomics-based or DL-based AI models to predict treatment response.

The literature search resulted in 347 total citations, with 135 from PubMed, 210 from Google Scholar, and 2 from further reviews (Figure 1). After removing duplicates and screening through the titles and abstracts, we screened 28 unique citations. During the full-text assessment phase, 7 articles were excluded as they did not meet the inclusion criteria. 21 articles were included in the final analysis.

Studies were grouped according to treatment regimen (chemotherapy alone vs chemotherapy plus targeted therapy), AI model type (radiomics-based ML, DL, or delta-radiomics), validation design (internal vs external), and reference standard (RECIST 1.1, histopathology, or survival-based response). A detailed summary of included studies is provided in Table 1.

Considerable methodological heterogeneity was observed across studies. 14 studies relied exclusively on internal validation strategies, including cross-validation and bootstrapping, while 5 studies used a form of temporal or internal-external validation. Only 2 studies employed independent external validation cohorts, and none achieved full TRIPOD-AI level 2b or 3 validation standards. The definition of treatment response also varied. 17 studies used RECIST 1.1 as the reference standard, 3 studies used histopathologic tumor regression grade, and 1 study defined response based on an overall survival threshold.

The pooled area AUC for all studies combined was 0.86 (95% CI 0.79 to 0.94) (Figures 2, 3), indicating excellent overall diagnostic performance. Among the 21 studies, 15 provided sufficient data for pooled sensitivity and specificity analysis. Sensitivity was 0.83 (95% CI 0.81 to 0.86) (Figure 4), demonstrating the model’s strength in correctly identifying responders. Specificity was 0.65 (95% CI 0.62 to 0.68) (Figure 5), suggesting moderate capability in correctly identifying nonresponders.

The findings of this systematic review and meta-analysis provide evidence that AI and radiomics-based models can play a significant role in predicting chemotherapy response in patients with CRLM. The pooled AUC of 0.86 for all studies supports the clinical utility of these tools in stratifying patients based on their likelihood of responding to systemic therapy.

Our findings are reinforced by multiple innovative radiomics strategies across included studies. Delta-radiomics ¹⁶ and dynamic feature analysis ¹⁸ offered superior early response prediction, aligning with the need for early identification of nonresponders to avoid unnecessary treatment and toxicity. ³⁵ Su ²⁰ and Yoon ²³ demonstrated the power of serial and longitudinal MRI, highlighting the importance of capturing temporal changes in tumor characteristics. ³⁶ Ensemble learning ²⁰ and attention-based DL ²⁵ showed performance advantages over traditional radiomics, supporting the impression that more sophisticated AI techniques can better extract and learn from complex imaging data. ³⁶ Integration of tumor microenvironment, ¹⁷ peritumoral features, ²¹ and biologic markers ²⁸ further improved prediction, underscoring the multifactorial nature of treatment response and the benefit of incorporating diverse data sources.

The particularly high sensitivity observed in the chemotherapy with targeted therapy subgroup is notable. This group included studies such as Zhou et al²⁸, who developed a multimodal model integrating PET/CT, clinical, and histopathological features (DERBY+), which achieved an AUC of 0.83. Similarly, Dercle et al demonstrated an AUC of 0.80 using a CT-based radiomics response signature in patients treated with anti-EGFR therapy. ²⁹ Yoon et al further validated the potential of MRI radiomics, achieving an AUC of 0.857 in a longitudinal DL model predicting early response to biologic therapy. ²³ Other key contributors to this subgroup include Zhu et al, ²⁴ Maaref et al, ³⁰ Miyamoto et al, ²² and Qu et al, ¹⁸ all of which underscore the value of integrating temporal and multiparametric imaging in biologic therapy response prediction. This finding may reflect the enhanced biological impact of agents like bevacizumab and cetuximab, which influence tumor vasculature and molecular expression profiles. These subtle changes are not always readily apparent through size-based RECIST criteria but may be detected through quantitative imaging biomarkers. AI models, especially those utilizing DL, are well-suited to capture these nuances due to their ability to learn hierarchical imaging features. ³⁷

In contrast, the chemotherapy-only subgroup showed slightly lower sensitivity and similar specificity. However, several high-performing studies still emerged within this group. For instance, Chen et al achieved an AUC of 0.818 by incorporating hepatobiliary phase MRI radiomics, ¹³ and Karagkounis et al reported an AUC of 0.87 with 100% sensitivity using post-treatment CT features from tumoral and peritumoral regions. ¹⁷ These findings illustrate that even without adjunct biologic therapy, advanced radiomics frameworks can deliver high diagnostic accuracy. Additional models such as those by Su et al, ²⁰ Rabe et al, ¹⁹ and Giannini et al ^15,16 further support the feasibility of using imaging-based AI approaches, particularly those leveraging longitudinal analysis, lesion-level heterogeneity, and texture-specific features, for predicting chemotherapy response with strong clinical relevance.

The differential performance between chemotherapy-only and biologic-inclusive cohorts suggests that radiomics may be particularly sensitive to biologically targeted treatment effects. This supports the notion that AI-derived imaging biomarkers can serve as noninvasive surrogates for molecular or histologic assessment. ³⁸ The high AUCs seen in biologic-treated cohorts ^28-30 also validate AI’s role in predicting response to agents with complex mechanisms beyond cytotoxicity. ³⁹

Radiomics-based ML and DL models demonstrated comparable overall discrimination, although DL models showed higher pooled sensitivity and lower specificity, indicating a tendency to favor responder classification at the expense of misclassifying nonresponders. Delta-radiomics models yielded the highest pooled AUC despite fewer included studies, maintaining a more balanced sensitivity-specificity profile. These findings suggest that models leveraging temporal or treatment-induced imaging changes may better reflect true biologic response than those relying solely on baseline tumor features. From a clinical standpoint, highly sensitive models may reduce the risk of prematurely halting effective therapy, whereas models with more balanced accuracy may better prevent continuation of ineffective regimens.

However, the observed performance must be interpreted in the context of nontrivial methodological limitations. A large majority of studies relied on internal validation only, with relatively few employing temporal or internal-external designs and very few using independent external cohorts. The attenuation in AUC observed across validation strata illustrates the risk of optimism bias when models are not tested on independent datasets. This pattern mirrors the broader AI and radiomics literature and emphasizes the need for robust, multicenter external validation and adherence to reporting standards such as TRIPOD-AI.

Outcome definitions also varied, with most studies using RECIST 1.1, some using histopathologic tumor regression, and one using survival-based endpoints. RECIST-based models dominated the pooled estimates but also carried the intrinsic limitations of size-based criteria, which may not fully reflect biologic response, especially in the setting of targeted therapies. Histopathology-based models showed numerically higher AUCs, suggesting that biologically grounded endpoints may yield more meaningful predictive signatures; however, they were fewer in number and often limited to resectable cohorts. Survival-based models are conceptually attractive but more challenging to standardize and were underrepresented in the current literature. Collectively, this heterogeneity in outcome definitions complicates direct comparison between studies. A shift toward histopathologic endpoints, survival metrics, or composite response criteria incorporating functional imaging (eg, DWI, PET-CT) would enable more clinically meaningful prediction models.

While the pooled sensitivity of 0.86 across all studies, and 0.88 within the biologics subgroup, is encouraging, clinical applicability depends not only on the ability to detect responders but also to accurately identify nonresponders. The observed specificity of approximately 0.65 suggests that a considerable proportion of patients may be misclassified as likely responders despite potentially limited benefit. This could lead to unnecessary exposure to chemotherapy-related toxicities or delay in transitioning to alternative therapies. One contributing factor may be the difficulty in distinguishing stable disease from partial response using imaging alone. These models, though highly sensitive to subtle image-derived changes, may lack access to molecular or serum biomarkers that could refine response prediction. Therefore, future models must strike a balance between sensitivity and specificity to optimize treatment planning. Integrating genomic data or circulating tumor markers such as CEA, KRAS/NRAS status, or ctDNA levels may improve the discriminatory power of AI models and reduce false-positive rates. ^40,41

Heterogeneity in imaging protocols, segmentation approaches, feature extraction pipelines, and model architectures further complicates generalizability. Most studies were single-center, used institution-specific imaging parameters, and applied heterogeneous feature engineering and selection strategies. This variability likely contributed to the statistical heterogeneity observed in pooled estimates and limits the immediate transferability of any single model into routine practice. Standardization efforts, such as harmonized acquisition protocols, robust feature harmonization pipelines, and transparent reporting of preprocessing steps, will be critical for translating radiomics models beyond the research setting.

For clinical integration, several practical barriers must be addressed. These include the computational requirements of DL models, lack of interpretability (“black-box” nature), and the need for radiomics software validation. ⁴² Explainable AI techniques, such as SHAP values, Grad-CAM, or attention heatmaps, may help bridge the gap between algorithm output and clinical trust. ⁴³ Additionally, institutions with limited access to advanced imaging modalities or GPU infrastructure may face challenges in model deployment. Cost-effectiveness studies and integration into clinical decision support systems are needed to evaluate real-world feasibility.

Despite these limitations, the present analysis confirms that AI and radiomics have strong potential to augment traditional imaging criteria in CRLM, particularly for early response assessment and treatment stratification. The consistent signal of high AUCs across diverse models and settings suggests that quantitative imaging biomarkers can capture relevant aspects of tumor biology and therapeutic effect that are not readily visible to the human eye. For institutions seeking to adopt such tools, the current evidence supports prioritizing models that incorporate longitudinal or multiphasic imaging, leverage DL or hybrid architectures, and have undergone at least some form of external validation.

Future research should focus on prospective, multicenter trials that incorporate harmonized imaging protocols, standardized radiomics workflows, and predefined validation plans. Studies that link radiomics signatures with molecular and genomic data, as well as with circulating biomarkers, may enable more comprehensive and biologically grounded predictive models. In parallel, development and testing of explainable AI frameworks and formal cost-effectiveness analyses will be necessary to clarify how these tools can be safely and efficiently integrated into real-world oncologic decision-making.

This meta-analysis affirms the promising diagnostic value of radiomics-based AI models in predicting chemotherapy response in CRLM. These models perform particularly well in patients receiving chemotherapy in combination with targeted therapy, reflecting the value of integrating biologically rich treatment data into AI frameworks. With continued refinement and validation, these technologies have the potential to revolutionize how oncologic imaging is interpreted and applied in treatment decision-making.