Impact of Radiomic and Artificial Intelligence on Colorectal Cancer: A Narrative Review

Abstract

Colorectal cancer (CRC) remains one of the leading causes of cancer-related morbidity and mortality worldwide, representing a major public health challenge. Despite advances in screening strategies, surgical techniques, and systemic therapies, patient prognosis is often compromised by late diagnosis, tumor heterogeneity, and therapeutic resistance. In recent years, the integration of advanced imaging analytics and artificial intelligence (AI) has opened new avenues for precision oncology. Radiomics, defined as the high-throughput extraction of quantitative features from medical images, has emerged as a promising tool to capture intratumoral heterogeneity and predict clinical outcomes in a non-invasive manner. When combined with AI, particularly machine learning and deep learning approaches, radiomics enables the development of predictive and prognostic models that may support treatment personalization. This narrative review provides a comprehensive overview of CRC epidemiology and risk factors, summarizes current diagnostic and clinical management strategies, and focuses extensively on radiomics and AI applications in CRC, including workflow standardization, feature extraction, clinical applications, and challenges for implementation in daily practice.

1. Introduction

Colorectal cancer (CRC) is the third most diagnosed malignancy and one of the leading causes of cancer-related death worldwide, with an estimated 1.93 million new cases and approximately 904,000 deaths reported in 2022 [1]. Despite a slight decline in mortality in high-income countries due to screening programs and improvements in therapy, CRC continues to pose a significant burden, particularly in low- and middle-income countries where access to prevention and treatment remains limited [2,3]. The development of CRC is strongly influenced by lifestyle and environmental factors, including diet, obesity, physical inactivity, alcohol consumption, and smoking, as well as genetic predisposition and inflammatory bowel disease [4,5,6,7]. Although colonoscopy and fecal immunochemical testing have significantly contributed to early detection, many patients are still diagnosed at advanced stages, when therapeutic options are limited and prognosis is poor [8,9]. Advances in molecular profiling have significantly improved the understanding of CRC biology, enabling the identification of actionable genetic alterations, including RAS and BRAF gene mutations and microsatellite instability (MSI) status, all of which are essential for guiding treatment selection [10,11]. However, tissue biopsy remains an invasive procedure that only provides information from a small portion of the tumor, limiting its ability to capture the spatial and temporal heterogeneity that characterizes CRC [12]. Medical imaging plays a central role in CRC management, particularly in staging, treatment planning, and follow-up. Conventional imaging modalities, including computed tomography (CT), magnetic resonance imaging (MRI), provide valuable morphological and functional information but often fail to capture the full complexity of tumor biology [13,14]. To overcome these limitations, radiomics has been introduced as an advanced imaging-based approach that quantitatively analyzes tumor features beyond visual assessment, as reported in Figure 1 [15,16]. Radiomics, particularly when combined with artificial intelligence (AI), offers the potential to generate non-invasive biomarkers capable of predicting treatment response, prognosis, and genetic profiles. These methods may substantially improve personalized treatment strategies and support clinical decision-making in CRC [17,18]. Despite the increasing number of reviews on AI and radiomics in CRC, most existing works either focus narrowly on specific clinical applications or provide broad overviews without critically appraising methodological rigor, model reproducibility, and real-world clinical readiness. This narrative review aims to provide a comprehensive synthesis of the epidemiological background and current diagnostic pathways of CRC, with a strong focus on the applications of radiomics and AI in this setting. We discuss the radiomic workflow, feature extraction, integration with machine learning and deep learning, clinical applications in rectal and colon cancer, as well as limitations and future perspectives for clinical implementation.

2. Data Source and Study Selection

In this narrative review, studies were selected from 2013 to 2025 using a meticulous screening method to examine the role of AI in the diagnosis and treatment of CRC, aiming to delineate the existing literature and identify gaps and domains of interest for potential investigation. The search involved an extensive examination of electronic databases, including ScienceDirect, PubMed, and Google Scholar, employing a blend of pertinent keywords such as “colorectal cancer”, “Artificial Intelligence”, “management”, “prognosis”, “treatment”, “diagnosis”, “monitoring”, and “prediction”, as well as subcategories of AI like “machine learning (ML)”, “deep learning (DL)”, and “natural language processing (NLP)” included within our search parameters. Inclusion criteria were original research articles or systematic/narrative reviews focusing on radiomics or AI applications in colorectal cancer; studies evaluating diagnostic, prognostic, predictive, or molecular characterization tasks; use of CT, MRI or multimodal imaging; availability of sufficient methodological detail regarding radiomic feature extraction, AI model development, or validation. Exclusion criteria included: conference abstracts, editorials, letters, or commentaries; studies not involving radiomics or AI; studies focusing exclusively on non-colorectal malignancies; preclinical, animal, or phantom studies; articles lacking key methodological information or outcome data. The authors subsequently reviewed all articles to eliminate those that were irrelevant. The search strategy aimed to present a critical synthesis of key insights from selected articles, reviewing the progress of AI techniques in CRC, and their future potential to influence diagnosis and management.

3. Epidemiology and Risk Factors

Diets high in red and processed meat, low fiber intake, obesity, and sedentary behavior are established contributors to CRC risk [19]. Adherence to a Mediterranean diet has been associated with a reduction in CRC risk, although it does not eliminate the possibility of developing the disease, as multiple genetic and environmental factors also play a role [20]. Alcohol consumption and smoking are also recognized risk factors [21]. Hereditary syndromes account for approximately 5% of CRC cases. Familial adenomatous polyposis (FAP) and Lynch syndrome significantly increase lifetime risk and require intensive surveillance [22,23]. Patients with long-standing inflammatory bowel disease (IBD), particularly ulcerative colitis, have an elevated CRC risk, linked to chronic inflammation and genomic instability [24]. Emerging evidence suggests that the gut microbiome plays a role in modulating carcinogenesis. Dysbiosis is linked to chronic inflammation, which is a major risk factor for CRC, but dysbiosis does not directly lead to hereditary syndromes like Lynch syndrome or familial adenomatous polyposis (FAP) [25,26,27]. It is important to mention the role of the gut microbiome in shaping local inflammatory and metabolic environments, further contributing to interpatient variability in tumor behavior. Patients with Lynch syndrome carry germline mutations in DNA mismatch repair (MMR) genes, predisposing them to early-onset CRC and extracolonic cancers [26]. FAP, caused by germline mutations in the APC gene, is characterized by hundreds to thousands of colorectal adenomas and a near-100% lifetime CRC risk if left untreated [25]. This epidemiologic heterogeneity supports the rationale for AI-based modeling, as radiomics and machine learning are uniquely positioned to quantify complex, multiscale phenotypic patterns that traditional risk stratification approaches cannot fully capture.

4. Diagnosis and Clinical Management

Early detection of CRC dramatically improves survival outcomes. Colonoscopy remains the gold standard for diagnosis, allowing direct visualization, biopsy, and removal of premalignant polyps. However, fecal immunochemical testing (FIT) and stool DNA testing represent effective non-invasive alternatives for screening [8,28,29]. Recent innovations include blood-based liquid biopsy assays targeting circulating tumor DNA, which are under investigation for CRC detection and monitoring [30]. Accurate staging is essential to guide therapy. The TNM classification system, provided by the American Joint Committee on Cancer (AJCC), remains the cornerstone for stratifying patients based on tumor invasion, nodal involvement, and distant metastases [31]. Computed tomography (CT) is widely used for staging, particularly for detecting distant metastases, whereas magnetic resonance imaging (MRI) is preferred for local staging of rectal cancer due to its superior soft-tissue resolution [32]. Also, positron emission tomography (PET), often combined with CT, provides complementary functional information on metabolic tumor activity and may be useful in the detection of occult metastases [33]. For localized disease, surgery remains the primary treatment. In rectal cancer, neoadjuvant chemoradiotherapy followed by total mesorectal excision (TME) represents the standard of care for locally advanced tumors, reducing the risk of local recurrence [34]. Adjuvant chemotherapy is recommended for stage III CRC and considered for high-risk stage II cases [35,36]. In metastatic disease, systemic therapy is based on fluoropyrimidines combined with oxaliplatin or irinotecan, often in combination with targeted agents such as anti-EGFR (cetuximab, panitumumab) or anti-VEGF (bevacizumab) monoclonal antibodies, depending on RAS/BRAF mutational status and microsatellite stability [37]. While these conventional imaging and molecular approaches are essential, they remain limited by qualitative interpretation, sampling bias, and inability to fully capture spatial intratumoral heterogeneity. Radiomics and AI directly address these gaps by extracting high-dimensional, quantitative features from standard CT, MRI, and PET images, enabling objective characterization of tumor phenotype. Unlike molecular profiling, which relies on localized tissue samples, radiomics provides a whole-tumor, non-invasive assessment that can reflect heterogeneity, treatment response, and early biological changes not visible to the human eye. AI-driven predictive models can integrate imaging, clinical, and molecular data to refine risk stratification, anticipate treatment response, and support personalized therapeutic decision-making, thereby enhancing existing workflows rather than replacing them. Despite advances in molecular characterization and therapeutic strategies, prognosis remains heterogeneous, with significant variability in treatment response. This highlights the urgent need for novel biomarkers and predictive tools, paving the way for the integration of radiomics and AI in CRC management.

5. Radiomics and Artificial Intelligence

Radiomics has emerged as a promising field that aims to overcome the limitations of conventional imaging by converting medical images into high-dimensional quantitative data, thereby capturing intratumoral heterogeneity in a reproducible and non-invasive manner [38,39]. First introduced by Lambin et al. in 2012, radiomics is based on the hypothesis that medical images contain hidden information beyond what is visible to the human eye, and that advanced computational analysis can reveal imaging biomarkers with diagnostic, prognostic, and predictive value [40]. By integrating radiomic features with genomic, proteomic, and clinical data, radiomics bridges the gap between imaging phenotypes and tumor biology, offering new opportunities for precision oncology [41]. The radiomics pipeline, often referred to as the “Radiomic Flow,” consists of a multi-step process encompassing image acquisition, preprocessing, segmentation, feature extraction, and data analysis [42].

5.1. Image Acquisition and Preprocessing

The accuracy of radiomics heavily depends on the quality and consistency of image acquisition. Computed tomography (CT), magnetic resonance imaging (MRI) are the most widely employed modalities, while ultrasound is occasionally used for superficial tumors and vascular assessment [43]. Variability in acquisition protocols across institutions represents a major challenge, as differences in slice thickness, contrast enhancement, and reconstruction kernels can significantly alter extracted radiomic features [44]. To mitigate this, initiatives such as the Quantitative Imaging Biomarker Alliance (QIBA) and the Quantitative Imaging Network (QIN) have promoted standardized acquisition protocols to enhance reproducibility [45]. Preprocessing techniques—including noise reduction, intensity normalization, histogram equalization, and spatial resampling—are applied to homogenize datasets and reduce artifacts [46]. While such steps improve comparability, they may also introduce bias, underscoring the need for transparent reporting and protocol standardization. Open-source tools such as PyRadiomics and 3D Slicer have facilitated reproducibility by providing standardized preprocessing pipelines and feature extraction libraries [47,48]. To contextualize segmentation variability, recent benchmarking initiatives such as the Medical Segmentation Decathlon (MSD) and Beyond the Cranial Vault (BCV) challenge have provided standardized datasets for evaluating algorithmic performance across different abdominal and pelvic structures, including bowel and soft-tissue regions relevant to colorectal imaging. These challenges consistently demonstrate that segmentation accuracy varies substantially across scanners, institutions, and anatomical sites, with rectal and pelvic organs among the most difficult to segment reliably. Their findings highlight why manual, semi-automated, and fully automated segmentation methods differ in consistency and why robust benchmarking is essential for interpreting variability in radiomics studies.

5.2. Segmentation of Regions of Interest

Precise tumor segmentation is a critical yet challenging step in radiomics, as tumor borders are often irregular or ill-defined. Manual segmentation, while still widely used, is prone to inter- and intra-observer variability. Semi-automated methods reduce subjectivity by combining operator input with computational algorithms, whereas fully automated approaches leverage artificial intelligence, particularly convolutional neural networks (CNNs), to achieve fast and reproducible delineation [49]. Volumetric segmentation (VOI) generally provides more robust data than two-dimensional approaches, as it captures the entire heterogeneity of the lesion [50]. Emerging approaches also exploit multiparametric imaging to define “habitats,” or physiologically distinct tumor subregions. For instance, combining T2-weighted and diffusion-weighted MRI can reveal areas of necrosis or high cellular density, enhancing biological interpretability [51].

5.3. Feature Extraction

Once segmentation is complete, quantitative features are extracted from the defined regions of interest. These features are typically divided into semantic and agnostic categories [52]. Semantic features include descriptors routinely used in clinical practice, such as tumor size, shape, and enhancement patterns, whereas agnostic features—often referred to as texture features—capture spatial relationships between pixels or voxels, quantifying heterogeneity beyond human perception. First-order statistics derived from image histograms measure intensity distributions, skewness, kurtosis, entropy, and energy. These descriptors reflect the degree of uniformity or randomness in voxel intensities, which can be associated with necrosis, fibrosis, or cellular proliferation [53]. Second-order texture features are derived from matrices such as the gray-level co-occurrence matrix (GLCM), gray-level run-length matrix (GLRLM), and neighborhood gray-tone difference matrix (NGTDM). These capture relationships between adjacent voxels and provide insights into tissue architecture and heterogeneity [54]. For example, high entropy values reflect disorganized architecture, while high contrast indicates sharp transitions between regions of different density. Higher-order features further explore structural complexity using advanced mathematical methods such as wavelet transforms, fractal analysis, and Minkowski functionals, which capture multi-scale patterns and morphological irregularities [55]. The use of these advanced descriptors has expanded the ability of radiomics to model tumor biology in a quantitative, non-invasive manner.

5.4. Data Analysis and Dimensionality Reduction

Radiomics typically produces hundreds to thousands of features, creating high-dimensional datasets where the number of variables often exceeds the number of patients. This so-called “curse of dimensionality” increases the risk of overfitting and reduces generalizability [56]. To address this, dimensionality reduction methods are applied. Feature selection strategies, including statistical tests and machine learning–based ranking, help identify the most informative variables, while feature extraction methods such as principal component analysis (PCA) create composite features that summarize information from correlated variables [57]. Among feature selection techniques, the least absolute shrinkage and selection operator (LASSO) regression is one of the most widely used, as it effectively reduces redundancy and enhances model interpretability [58]. Balanced sampling approaches, such as the synthetic minority oversampling technique (SMOTE), are also employed to address class imbalance in training datasets [59]. Validation is essential to ensure clinical applicability. Techniques such as k-fold cross-validation, bootstrapping, and nested validation frameworks are recommended to evaluate robustness and avoid optimistic bias. Ultimately, external validation in independent multicenter datasets remains the gold standard, though it is rarely achieved in current radiomics studies [60]. The complexity and high dimensionality of radiomic datasets have accelerated the integration of artificial intelligence (AI) methods, particularly machine learning (ML) and deep learning (DL), into radiomics workflows. These approaches can handle vast numbers of features, identifying complex, non-linear relationships, and improving predictive accuracy beyond what is achievable with traditional statistical modeling [61]. Machine learning methods commonly applied in radiomics include supervised classifiers such as logistic regression, support vector machines (SVM), random forests, k-nearest neighbors (k-NN), and naïve Bayes [62]. These algorithms map extracted features to clinical outcomes, such as treatment response or overall survival. By classical machine learning (ML), we refer to traditional supervised learning algorithms that rely on engineered features rather than end-to-end deep learning. Examples include support vector machines (SVM), random forests (RF), logistic regression, k-nearest neighbors (k-NN), and gradient boosting methods such as XGBoost (version 3.1.2.1). These models typically operate on handcrafted radiomic features and have been widely used in early radiomics studies for prediction and classification tasks. Random forests, for example, are particularly effective at handling heterogeneous datasets, while SVMs have shown promise in small-to-medium sample sizes by constructing optimal hyperplanes for classification tasks [63]. Among feature selection methods, LASSO regression has been particularly influential for radiomics because it simultaneously performs feature reduction and predictive modeling, improving interpretability while avoiding overfitting [64]. Deep learning approaches, especially convolutional neural networks (CNNs), have further revolutionized radiomics. Unlike classical ML, which relies on hand-crafted features, CNNs automatically learn hierarchical features directly from raw imaging data. CNN-based segmentation has achieved remarkable performance in delineating tumors and surrounding tissues, often outperforming manual or semi-automated methods [65]. Furthermore, interpretability is being enhanced by explainable AI methods, such as Gradient-weighted Class Activation Mapping (Grad-CAM), which visualizes the most influential regions of the image contributing to predictions [66].

6. Clinical Applications of Radiomics in Colorectal Cancer

6.1. Prediction of Therapy Response

In locally advanced rectal cancer (LARC), radiomics has shown promise in predicting response to neoadjuvant chemoradiotherapy (nCRT). Radiomic features derived from baseline T2-weighted MRI and diffusion-weighted imaging (DWI) have been associated with complete pathological response, with SVM and random forest models achieving area under the ROC curve (AUC) values between 0.71 and 0.87 [67,68]. For example, He J et al. [69] developed an MRI-based radiomics and machine-learning model using pretreatment multiparametric MRI to predict pathological complete response (pCR) in patients with locally advanced rectal cancer. Their model demonstrated high discriminative performance (AUC > 0.85) and was explicitly proposed as a tool to identify potential candidates for organ-preserving approaches. In addition, Domingo et al. developed a transcriptomic ML model to predict pCR in rectal cancer treated with nCRT. The authors emphasized the role of such predictive tools in refining patient selection for organ preservation, providing biological insights into therapy response [70]. In a recent study radiomic features extracted from restaging MRI were able to distinguish good responders from non-responders, improving discrimination beyond conventional radiological criteria. This work demonstrated that texture-based signatures can capture subtle microstructural changes induced by therapy and may support more accurate identification of patients eligible for non-operative management [71].

6.2. Assessment of Vascular and Perineural Invasion

Extramural venous invasion (EMVI) and perineural invasion (PNI) are critical prognostic factors in rectal cancer, associated with higher risks of recurrence and poor survival. Conventional MRI often fails to detect these features preoperatively due to edema or fibrosis. Radiomic models trained on T2-weighted MRI have demonstrated superior accuracy compared to radiologists in identifying EMVI and have shown promise in predicting PNI, supporting more individualized treatment strategies [72,73].

6.3. Prediction of Liver Metastases

Both synchronous and metachronous liver metastases (SLM and MLM) represent major clinical challenges in colorectal cancer. Radiomics nomograms combining clinical factors and radiomic signatures from T2-weighted MRI have achieved excellent predictive performance for synchronous liver metastases, with AUCs as high as 0.94 [74,75]. For metachronous metastases, logistic regression models integrating radiomic features from baseline rectal MRI have provided robust risk stratification, highlighting the role of non-invasive imaging biomarkers in early detection [76].

6.4. Prediction of Genetic Mutations

Molecular profiling is crucial for guiding targeted therapies, particularly for mutations in KRAS, NRAS, and BRAF, which influence the efficacy of anti-EGFR agents. Radiomics has emerged as a non-invasive alternative to biopsy, with multiple studies demonstrating significant associations between radiomic features and mutational status. MRI-based texture analysis has differentiated KRAS mutation status with accuracies exceeding 80%, while multicenter CT-based models using SVM classifiers have predicted KRAS and BRAF mutations with AUCs approaching 0.80 [77,78,79]. Radiomic features were shown to correlate with transcriptomic programs and clinically relevant phenotypes, revealing that specific imaging patterns reflect distinct molecular pathways. This integrative framework provides mechanistic support for the use of radiomics as a surrogate biomarker and demonstrates how the combination of imaging and molecular data can enhance prognostic stratification and biological interpretability [80]. These findings underscore the potential of radiomics as a surrogate biomarker for molecular characterization.

6.5. Delta-Radiomics

An emerging frontier is delta-radiomics, which evaluates temporal changes in radiomic features before and after therapy. This approach provides dynamic insights into tumor biology, offering prognostic information beyond static baseline imaging [81]. In colorectal liver metastases, delta-radiomic signatures have predicted response to chemotherapy and anti-EGFR therapies more accurately than RECIST criteria, achieving AUCs of 0.80 or higher [82,83,84]. By quantifying intralesional heterogeneity and treatment-induced changes, delta-radiomics could enable real-time adaptation of therapeutic strategies (

Table 1. Summary of key studies on the application of radiomics and artificial intelligence for predicting clinical outcomes in colorectal cancer.

Author/s and Year (Ref.)	Clinical Application	Imaging Modality	AI/Radiomics Methodology	Main Outcome (Metrics)
Liu Z, et al. (2017) [83]	Prediction of pCR to nCRT	MRI (T2WI + DWI)	Radiomics model (LASSO + SVM + clinical)	AUC = 0.98 (in the external validation cohort)
Cui Y, et al. (2019) [84]	Prediction of pCR to nCRT	MP-MRI (T2w, cT1w, ADC)	Radiomics nomogram (LASSO + clinical)	AUC = 0.966 (in the validation cohort)
Horvat N, et al. (2018) [68]	Prediction of pCR to nCRT	MRI (T2-weighted)	Radiomics score (LASSO)	AUC = 0.77 (in the external validation cohort)
Zhang YC, et al. (2022) [85]	Differentiating tumor deposits (TDs) from lymph node metastasis (LNM)	CT	Radiomics signature from the largest peritumoral nodule	AUC = 0.918 (in the validation cohort)
Shu Z, et al. (2022) [86]	Prediction of extramural venous invasion (EMVI)	MP-MRI	Joint model (Bayes-based radiomics + clinical factors)	AUC = 0.835 (in the test set)
Chen J, et al. (2021) [73]	Prediction of Perineural Invasion (PNI)	MRI (T2-weighted)	Radiomics nomogram (mRMR & LASSO)	AUC = 0.85 (in the test cohort)
Bae JS, et al. (2019) [72]	Assessment of Extramural Venous Invasion (EMVI)	MRI (T2-weighted)	Radiomics assessment (ROC curve analysis)	AUC = 0.829 (experienced radiologist)
Oh JE, et al. (2020) [79]	Differentiation of KRAS mutation status	MRI (T2-weighted)	Texture analysis (Decision tree model)	AUC = 0.884 (on the whole dataset)
Meng X, et al. (2019) [77]	Prediction of various biological characteristics	MP-MRI	Radiomics signature (various selectors & classifiers)	AUCs (validation): Differentiation = 0.720; Ki-67 = 0.699 KRAS = 0.651
Liu M, et al. (2020) [75]	Prediction of synchronous liver metastasis (SLM)	MRI (T2-weighted)	Radiomics nomogram (LASSO + clinical factors)	AUC = 0.944 (in the validation cohort)
Giannini V, et al. (2022) [81]	Prediction of therapy response of liver metastases to FOLFOX	CT	Delta-radiomics signature (Decision tree)	AUC = 0.93(in the validation cohort)
Shu Z, et al. (2019) [74]	Prediction of synchronous liver metastasis (SLM)	MRI (T2-weighted)	Radiomics nomogram (feature selection with LASSO)	AUC = 0.912 (in the validation cohort)
Dercle L, et al. (2020) [80]	Prediction of therapy response to anti-EGFR treatment	CT	Delta-radiomics signature (Random Forest)	AUC = 0.80 (in the validation cohort)
Chuanji Z, et al. (2022) [87]	Prediction of 3-year overall survival (OS)	MRI (T2-weighted)	Comprehensive nomogram (radiomics, morphological & clinical)	C-index = 0.944 (in the validation cohort)

Abbreviations: pCR, Pathologic Complete Response; nCRT, neoadjuvant Chemoradiotherapy; MRI, Magnetic Resonance Imaging; T2WI, T2-weighted Imaging; DWI, Diffusion-weighted Imaging; MP-MRI, Multiparametric MRI; cT1w, contrast-enhanced T1-weighted; ADC, Apparent Diffusion Coefficient; CT, Computed Tomography; SVM, Support Vector Machine; LASSO, Least Absolute Shrinkage and Selection Operator; EMVI, Extramural Venous Invasion; PNI, Perineural Invasion; mRMR, Maximum Relevance Minimum Redundancy; TD, Tumor Deposits; LNM, Lymph Node Metastasis; KRAS, Kirsten Rat Sarcoma viral oncogene homolog; EGFR, Epidermal Growth Factor Receptor; FOLFOX, Folic acid + Fluorouracil + Oxaliplatin; OS, Overall Survival; C-index, Concordance Index.

). An emerging frontier is delta-radiomics, which evaluates temporal changes in radiomic features before and after therapy. This approach provides dynamic insights into tumor biology, offering prognostic information beyond static baseline imaging [85]. In colorectal liver metastases, delta-radiomic signatures have predicted response to chemotherapy and anti-EGFR therapies more accurately than RECIST criteria, achieving AUCs of 0.80 or higher [86,87]. By quantifying intralesional heterogeneity and treatment-induced changes, delta-radiomics could enable real-time adaptation of therapeutic strategies (Table 1). However, several limitations currently restrict clinical applicability. The reproducibility of longitudinal imaging is a major challenge: variations in patient positioning, bowel motion, contrast timing, or scanner calibration across different time points can introduce artificial fluctuations in radiomic features, potentially confounding true biological changes. Moreover, acquisition parameters often differ between baseline and follow-up scans—such as slice thickness, reconstruction kernels, or diffusion-weighted imaging settings—leading to instability in temporal feature measurements. Another critical issue is the lack of temporal standardization: the timing of post-treatment imaging varies widely across studies and clinical workflows, making it difficult to disentangle genuine treatment effects from differences in imaging schedules. Finally, the absence of large, multi-institutional test–retest datasets limit the ability to assess longitudinal feature robustness. Addressing these challenges through harmonized acquisition protocols and validated temporal normalization strategies will be essential for the reliable clinical adoption of delta-radiomics.

7. Limitations and Current Challenges

Despite its rapid growth, radiomics remains largely confined to research settings and has not yet been fully integrated into clinical practice (Figure 2). A key challenge lies in the variability of imaging protocols across institutions. Differences in acquisition parameters, reconstruction algorithms, and segmentation methods result in features that are not always reproducible, which undermine the reliability of predictive models [88]. Manual segmentation, still widely used, introduces inter- and intra-observer variability, while even automated methods can be affected by the quality and diversity of the training datasets [89]. A further limitation relates to the clinical representativeness of current radiomic datasets. Most models are trained on relatively homogeneous cohorts, often excluding patients with significant comorbidities or atypical disease presentations. As a result, models may perform well within the restricted distribution of the training data but fail to generalize when applied to patients with underlying conditions—such as inflammatory disorders, prior surgeries, metabolic disease, or treatment-related tissue changes—that alter imaging appearance but are underrepresented in training samples. This mismatch increases the risk of erroneous predictions when models are deployed in real-world populations [90]. Similarly, demographic bias is a growing concern. Radiomic studies frequently rely on single-center cohorts with narrow demographic profiles, including limited variability in age, ethnicity, socioeconomic status, and genetic background. Although disease progression mechanisms may be shared, the biological and imaging phenotypes of colorectal cancer (CRC) can differ across populations, and factors such as body habitus, tumor biology, or access to healthcare may influence both imaging signatures and treatment response. These constraints reduce model scalability and challenge their translation into diverse clinical environments. Another limitation is the lack of standardization in feature definitions. The same features may be calculated differently depending on the software or preprocessing pipeline, leading to inconsistencies across studies. Initiatives such as the Image Biomarker Standardization Initiative (IBSI) are working to address these discrepancies, but widespread adoption is still limited [91]. Moreover, most studies are retrospective, single-center analyses with relatively small sample sizes. These methodological constraints limit generalizability, making external validation across diverse patient cohorts imperative before radiomic biomarkers can be translated into routine care [92]. Ethical and regulatory considerations are increasingly central to AI-driven oncology. Future implementations must ensure robust patient consent, rigorous data anonymization and de-identification procedures, and transparent documentation of algorithmic decision-making. Emerging regulatory frameworks from the FDA and EMA further highlight the need for explainable AI systems, as model interpretability is essential to guarantee clinical accountability and foster clinician and patient trust. Finally, ethical and regulatory frameworks for AI-driven decision support in oncology are still evolving, and issues of data privacy, model transparency, demographic fairness, and clinical accountability must be resolved before radiomics-based tools can be responsibly integrated into standard workflows [93].

8. Future Perspectives

Radiomics and AI have the potential to transform the management of colorectal cancer by enabling non-invasive, quantitative, and individualized assessments of tumor biology. Moving forward, several priorities can accelerate clinical translation. First, standardization of imaging acquisition, segmentation, and feature extraction is essential to ensure reproducibility. Collaborative multicenter studies and the establishment of large, harmonized databases will provide the statistical power and diversity necessary for robust external validation [94]. Second, integration of radiomics with other “omics” technologies—such as genomics, transcriptomics, and proteomics—offers the promise of multi-modal predictive models that capture complementary aspects of tumor biology, providing deeper insights into mechanisms of progression and resistance [95]. Third, the development of explainable AI (XAI) frameworks is crucial to improving the interpretability of predictive models. Transparent, human-interpretable algorithms will facilitate clinician trust and accelerate adoption into multidisciplinary tumor boards [96]. In parallel, generative AI (GenAI) emerges as a powerful tool to address current limitations in data availability and heterogeneity. It can generate realistic, high-fidelity cell-level images that preserve key morphological phenotypes of cancer stem cells. Such approaches provide a proof-of-concept for using GenAI to augment datasets, model tumor heterogeneity, and simulate rare or underrepresented biological patterns. Applied to colorectal cancer radiomics, similar generative frameworks could improve model robustness, support domain adaptation across imaging centers, and ultimately enhance the biological relevance of predictive models [97]. Finally, prospective clinical trials incorporating radiomics-guided decision-making are needed to establish clinical utility. These trials could evaluate whether radiomics-based biomarkers improve patient outcomes by refining treatment selection, reducing overtreatment, and personalizing follow-up strategies [46]. To accelerate real-world implementation, several actionable steps should be prioritized. Establishing a dedicated CRC radiomics consortium would facilitate standardized imaging protocols, multicenter data harmonization, and cross-institutional reproducibility studies. Integrating radiomic signatures into molecular tumor boards could support multidisciplinary decision-making by combining imaging-derived phenotypes with genomic and pathological data. In addition, developing shared CRC radiomics repositories and benchmarking datasets would enable transparent evaluation of model performance across institutions. Finally, prospective clinical trials explicitly incorporating radiomics-guided stratification are needed to determine whether these tools can measurably improve treatment selection, escalate or de-escalate therapy, and enhance patient outcomes.

9. Conclusions

Radiomics and artificial intelligence represent a paradigm shift in the evaluation and management of colorectal cancer. By extracting high-dimensional, quantitative features from standard medical imaging, radiomics provides novel insights into tumor heterogeneity, treatment response, and molecular status, offering non-invasive biomarkers for precision oncology. The integration of AI, particularly machine learning and deep learning, enhances the predictive power of radiomics, enabling dynamic tools such as delta-radiomics that monitor treatment-induced changes over time [98,99]. However, significant barriers remain, including a lack of standardization, limited external validation, and concerns regarding interpretability. Addressing these challenges through multicenter collaborations, standardized workflows, and explainable AI will be critical for clinical implementation. If successfully translated, radiomics and AI hold the potential to complement histopathology and molecular profiling, advancing precision medicine and improving outcomes in colorectal cancer patients.