Expanding the role of radiomics and artificial intelligence in the management of inflammatory bowel disease: Insights, opportunities, and challenges

Abstract

Inflammatory bowel disease (IBD), encompassing Crohn’s disease and ulcerative colitis, remains a chronic management challenge despite the success of biological therapies such as infliximab. A major limitation is secondary loss of response, which affects a substantial proportion of patients and complicates long-term treatment strategies. Emerging technologies such as radiomics, which converts medical images into quantitative features, and artificial intelligence (AI), which integrates complex multimodal data, offer new opportunities to predict treatment response, monitor disease activity, and personalize therapy. By combining imaging-derived radiomic features with clinical and laboratory information, AI-driven models can provide early, actionable insights to guide therapeutic decisions. This editorial discusses the promise and limitations of these approaches, emphasizing how they can be integrated into clinical decision-making pathways. While challenges in standardization, validation, and clinician adoption remain, radiomics and AI represent important steps toward precision medicine, with the potential to improve outcomes and optimize care for patients with IBD.

Key Words: Radiomics; Artificial intelligence; Inflammatory bowel disease; Crohn’s disease; Ulcerative colitis; Secondary loss of response; Infliximab; Precision medicine; Machine learning; Disease monitoring

Core Tip: Radiomics and artificial intelligence are emerging as powerful tools in the management of inflammatory bowel disease (IBD), particularly in predicting secondary loss of response to biologic therapies like infliximab. By integrating imaging data with clinical and laboratory parameters, these technologies offer a more personalized, proactive approach to disease management, enabling earlier interventions and optimizing treatment plans for Crohn’s disease and ulcerative colitis. While challenges remain in standardization and clinical adoption, the potential to transform IBD care with precision medicine is significant.

INTRODUCTION

Inflammatory bowel disease (IBD), comprising Crohn’s disease (CD) and ulcerative colitis (UC), is a chronic, relapsing inflammatory disorder of the gastrointestinal tract affecting > 7 million people worldwide, with rising incidence in both developed and developing regions[1,2]. It often presents in early adulthood and imposes a lifelong burden on quality of life, productivity, and healthcare systems. Although they share overlapping symptoms, CD and UC differ pathologically. CD can involve any part of the gastrointestinal tract with discontinuous, transmural inflammation leading to strictures and fistulas, whereas UC is limited to the colon and rectum with continuous, mucosal inflammation[1,2].

The introduction of biological therapies, particularly infliximab (IFX), has transformed IBD management. By targeting tumor necrosis factor-α, IFX induces and maintains remission, promotes mucosal healing, and reduces the need for surgery and hospitalization[3]. Yet long-term disease control remains elusive. A major limitation is secondary loss of response (SLOR), whereby patients who initially respond to IFX lose efficacy over time due to factors such as immunogenicity, low trough levels, evolving disease biology, or persistent inflammatory burden[4]. SLOR is common, affecting 20%–50% of patients with CD and 15%–30% with UC within the first year, and is associated with relapse, treatment escalation or switching, repeated investigations, and significant costs[5-7].

Conventional monitoring, clinical scores, biomarkers such as C-reactive protein and fecal calprotectin, therapeutic drug monitoring, and endoscopy provide valuable information but are often reactive, invasive, or insufficiently predictive of individual outcomes. The European Crohn’s and Colitis Organization guidelines advocate for proactive, treat-to-target strategies with objective disease assessment[8]. However, current tools may not deliver the precision needed to anticipate SLOR and personalize therapy.

Radiomics and artificial intelligence (AI) offer new opportunities. Radiomics converts routine imaging into quantitative features that capture tissue heterogeneity and subtle changes invisible to the human eye[8,9]. AI, including machine learning (ML) and deep learning (DL), can integrate these imaging features with clinical, laboratory, and pharmacokinetic data to generate predictive models[10,11]. Such models could identify patients at risk of SLOR, forecast treatment response, and support earlier, proactive intervention[12]. This editorial examines the potential of radiomics and AI to reshape IBD management. We first consider radiomics and its application in CD and UC, then explore how AI integrates imaging and clinical data to advance precision medicine. Finally, we discuss challenges, ethical considerations, and future directions, highlighting how these technologies may shift care from reactive to proactive, personalized strategies.

RADIOMICS IN IBD MANAGEMENT

Radiomics is founded on the principle that medical images contain a lot of information beyond that perceived by the human eye. While conventional radiological interpretation focuses on descriptive features such as bowel wall thickness, mural enhancement, or mesenteric changes, radiomics enables systematic quantification of hundreds or thousands of features that describe the intensity, shape, texture, and spatial relationships of tissues. These features, once extracted, can be correlated with biological processes, disease behavior, and treatment outcomes.

The radiomics pipeline typically includes image acquisition, segmentation of regions of interest, preprocessing, feature extraction, and statistical or ML analysis. Standardized imaging protocols are crucial, as variability in acquisition can alter feature reproducibility. Segmentation can be manual, semi-automated, or automated, with reproducibility analyses essential to ensure reliability. Preprocessing may include intensity normalization, resampling, and filtering to reduce artifacts. Feature extraction is performed using established libraries, often following guidelines such as those of the Image Biomarker Standardization Initiative[13,14]. In the context of IBD, radiomics has shown particular promise with imaging modalities such as computed tomography enterography (CTE), magnetic resonance enterography (MRE), and ultrasound. CTE provides excellent spatial resolution, allowing assessment of bowel wall enhancement, fat stranding, and mesenteric vascularity[15,16]. However, repeated radiation exposure is a concern, particularly in young patients requiring long-term monitoring. MRE has become the modality of choice for many centers, offering superior soft tissue contrast and functional sequences such as diffusion-weighted imaging and dynamic contrast-enhanced imaging. MRE can evaluate both inflammatory and fibrotic components, making it highly suitable for radiomic analysis[17-19]. Ultrasound, especially with contrast enhancement and elastography, is increasingly recognized as a valuable, noninvasive tool for assessing disease activity, though variability in operator performance and segmentation remains a challenge.

CD has been the primary focus of radiomics applications, given its transmural nature and heterogeneous disease patterns. Radiomics enables detailed characterization of bowel wall thickening, mural edema, ulcerations, and surrounding mesenteric changes. Creeping fat, the hypertrophied mesenteric fat enveloping inflamed bowel loops, is a particularly distinctive feature of CD. Quantitative radiomic analysis of creeping fat density and texture has shown correlations with inflammation, fibrosis, and adverse outcomes[20-22]. Features such as entropy and contrast derived from texture matrices, or skewness and kurtosis from intensity histograms, have been used to distinguish inflammatory from fibrotic lesions[22]. This distinction is clinically important, as inflammatory lesions may respond to medical therapy, while fibrotic strictures often require surgical intervention. In UC, radiomics applications are less well developed but equally promising. UC is confined to the mucosa and submucosa, making radiomic analysis of colonic wall thickness, vascularity, and texture particularly relevant. Studies suggest that radiomics can provide objective measures of mucosal inflammation, potentially complementing or substituting for endoscopy in certain scenarios[23,24]. Radiomic signatures may also help stratify patients according to risk of relapse, thereby informing treatment strategies.

One of the most compelling applications of radiomics is its potential to predict treatment response and SLOR[25]. Traditional monitoring methods typically detect treatment failure after it occurs, whereas radiomics may detect subtle changes that precede overt relapse. For example, increased bowel wall thickness, altered mural enhancement, or changes in texture heterogeneity on follow-up imaging have been associated with reduced IFX efficacy[26,27]. By comparing baseline and post-treatment imaging, radiomics can identify features predictive of treatment success or failure, allowing clinicians to adjust therapy proactively[28,29]. Radiomics also offers opportunities to develop imaging biomarkers that can be integrated into clinical trials and practice. Quantitative radiomic features could serve as surrogate endpoints, complementing endoscopic or histological outcomes. Moreover, radiomics may support stratification of patients in trials, enabling enrichment of populations most likely to benefit from particular interventions. As radiomics matures, its role in precision medicine is likely to expand, providing objective, reproducible measures of disease biology and therapeutic response.

AI AND PRECISION MEDICINE IN IBD

While radiomics provides a rich source of quantitative imaging features, AI is essential for unlocking its clinical potential. AI encompasses a range of computational approaches that enable machines to learn patterns from data and make predictions. In IBD, AI can integrate radiomic features with clinical, laboratory, and molecular data to generate predictive models of treatment response, disease progression, and SLOR. ML methods such as logistic regression with regularization, random forests, gradient boosting, and support vector machines have been applied to radiomic datasets with promising results[30,31]. These models can handle high-dimensional data, identify the most informative features, and generate predictions with good accuracy. Importantly, feature selection methods such as least absolute shrinkage and selection operator regression help prevent overfitting by isolating the most predictive variables. DL, particularly convolutional neural networks (CNNs), has further advanced AI applications in medical imaging. CNNs can automatically learn hierarchical representations from raw imaging data, capturing complex features of bowel wall thickening, mucosal enhancement, and mesenteric changes. Unlike traditional radiomics, which relies on handcrafted features, CNNs learn directly from data, enabling the discovery of novel imaging biomarkers. Hybrid approaches that combine handcrafted radiomic features with CNN-derived embeddings may offer the best of both worlds, balancing interpretability with predictive power[32].

The strength of AI lies not only in analyzing radiomic data but also in integrating multimodal information (Table 1). Predictive models can combine radiomic features with laboratory biomarkers such as C-reactive protein and fecal calprotectin, pharmacokinetic measures such as IFX trough levels and antidrug antibodies, and patient-level variables such as age, disease duration, and phenotype[33]. This multimodal integration provides a more comprehensive view of disease biology, enabling models to generate patient-specific risk profiles for remission, relapse, or SLOR[34,35]. Emerging approaches such as federated learning enhance AI applications in IBD. Federated learning enables models to be trained across multiple centers without sharing raw patient data, addressing privacy concerns and enhancing generalizability[36]. This is particularly important in IBD, where disease heterogeneity and variability in imaging protocols limit the robustness of single-center models. Federated learning allows multi-institutional collaboration while maintaining patient confidentiality, a critical step for real-world implementation.

Table 1 Key studies on radiomics, artificial intelligence, digital health, and predictive models in inflammatory bowel disease.

Refs	Study focus	Methodology/cohort	Key findings
Li et al[33], 2025	Radiomics & SLOR to infliximab in CD	220 CD patients, 2-center, retrospective; intestinal wall + creeping fat radiomics; combined clinical-radiomics model	Combined model achieved best prediction of SLOR (AUC: 0.87 training, 0.85 validation); outperforming clinical or radiomics alone
Cai et al[15], 2024	Fusion models for predicting IFX response in CD	263 CD patients, 3-center retrospective; CTE-based radiomics, DL, clinical, and fusion models	Early fusion (radiomics + DL + clinical) had highest accuracy (AUC: 0.85–0.91) and best calibration; outperformed single-modality models
Chirra et al[17], 2024	Detecting inflammation & fibrosis in stricturing CD	51 CD patients undergoing MRE before resection; radiomics vs radiologist vs histopathology	Radiomics differentiated inflammation vs fibrosis (AUC: 0.67–0.83); combined radiomics + radiologist scoring improved accuracy (AUC: Approximately 0.79)
Colombel et al[46], 2017	Long-term safety of vedolizumab in IBD	Pooled safety data: 2830 patients, > 4800 patient-years exposure	Vedolizumab safe long-term; low rates of serious infection (≤ 0.6%), infusion reactions, and malignancy (< 1%). No PML reported
Kennedy et al[28], 2019	Predictors of anti-TNF failure (PANTS study)	1610 anti-TNF-naïve CD patients, prospective UK-wide; IFX & ADA	Low week-14 drug levels predicted nonresponse and nonremission; immunogenicity frequent (63% IFX, 29% ADA); immunomodulators reduced antibody risk
Liu et al[18], 2024	ML diagnosis of ileal CD (radiomics + clinical)	135 patients, T2-weighted MRE radiomics; compared to radiologists	Radiomics model (AUC: 0.95) outperformed 2/3 radiologists; ensemble radiomics + clinical model highest performance (AUC: 0.98, 93.5% accuracy)
Maccioni et al[49], 2000	MRI correlation with CD activity	20 CD patients, 1.5T MRI vs biomarkers	MRI signals (wall T2, gadolinium enhancement, fibrofatty proliferation) strongly correlated with biologic activity, even in clinical remission
Peters et al[45], 2023	Transcriptomic temporal signatures in IBD	Mouse colitis models + IBD patient validation; ML-based temporal classifier	Temporal gene expression & splicing signatures predicted histopathology, distinguishing acute vs chronic phases; translational potential in IBD
Qiu et al[44], 2025	Predicting long-term IFX response	746 CD patients, multicenter; ML models (XGBoost, SHAP, LCMM)	XGBoost best (AUC: 0.91 train, 0.71 test); key predictors: Hb, WBC, ESR, albumin, platelets; identified distinct patient clusters
Rimola et al[29], 2011	MRI activity index validation	48 CD patients; MRI vs ileocolonoscopy (CDEIS)	MRI activity index validation strongly correlated with CDEIS (r = 0.80); validated thresholds for active (≥ 7) and severe (> 11) disease with high sensitivity/specificity
Song et al[16], 2024	CTE radiomics + body composition to predict IFX failure	137 CD patients; CTE + muscle/fat indices; 1-year IFX outcomes	Combined radiomics + skeletal muscle index + creeping fat model best (AUC: 0.88 train, 0.83 validation); strong clinical utility
Waljee et al[50], 2017	ML to predict hospitalization/steroid use	20368 VHA IBD patients; logistic regression vs random forest	RF model outperformed logistic regression (AUC: 0.87 vs 0.68); key predictors: Age, albumin, immunosuppressive use, platelets, prior hospitalizations/steroid use
Yang et al[12], 2024	Pancreatic radiomics to predict SLOR to IFX	184 biologic-naïve CD patients; pancreatic texture + clinical model	Combined clinical–radiomics nomogram best (AUC: 0.87); first pancreatic-based model for SLOR prediction
Yueying et al[19], 2023	MRE-based model for IFX response	188 bio-naïve CD patients; pretreatment MRE radiomics	Radiomic model achieved AUC: 0.88; robust across centers/scanners; high-risk group more likely to lose response
Zhang et al[21], 2024	MRI radiomics for intestinal fibrosis	123 refractory CD patients; MR-based fibrosis models vs radiologists	Radiomics models (AUC ≤ 0.93) outperformed visual reads and clinical markers; enhanced fibrosis characterization
Zhen et al[48], 2021	Digital health monitoring in IBD	59 IBD patients using HealthPROMISE app for 1 year	ER visits/hospitalizations reduced from 25% to 3% (P = 0.03); improved patient understanding of disease; engagement limited (~54%)
Zhu et al[27], 2022	CTE radiomics nomogram for mucosal healing	106 CD patients on IFX; training + test cohorts	Clinical–radiomics nomogram predicted MH (AUC: 0.88 both cohorts); suggested utility for noninvasive MH monitoring
Zhu et al[26], 2024	DECT radiomics for mucosal healing	106 CD patients, 221 segments; iodine maps + mesenteric fat	Combined iodine radiomics model best (AUC: 0.99 train, 0.95 test); high-risk patients had more progression
Zhu et al[25], 2023	CTE radiomics nomogram for SLOR to IFX	181 CD patients; multicenter; 3 cohorts	Radiomics nomogram predicted SLOR with high accuracy (AUC: 0.86-0.95 across cohorts); supported early biologic switching
Aggeletopoulou et al[20], 2023	Review of creeping fat in CD	Review of surgical, histologic, radiologic approaches	Creeping fat strongly linked to CD outcomes; novel CT-based creeping fat index proposed; ML + radiomics may clarify mechanistic role

SLOR: Secondary loss of response; CD: Crohn’s disease; AUC: Area under the curve; DL: Deep learning; MRE: Magnetic resonance enterography; IBD: Inflammatory bowel disease; PML: Progressive multifocal leukoencephalopathy; TNF: Tumor necrosis factor; PANTS: Personalised anti-tumor necrosis factor therapy in Crohn’s disease; IFX: Infliximab; ADA: Antidrug antibodies; ML: Machine learning; MRI: Magnetic resonance imaging; XGBoost: Extreme gradient boosting; SHAP: Shapley additive explanations; LCMM: Latent class mixed model; Hb: Hemoglobin; WBC: White blood cell; ESR: Erythrocyte sedimentation rate; CDEIS: Crohn’s Disease Endoscopic Index of Severity; CTE: Computed tomography enterography; VHA: Veterans Health Administration; RF: Random forest; MR: Magnetic resonance; ER: Emergency room; MH: Mucosal healing; DECT: Dual-energy computed tomography; CT: Computed tomography.

The application of AI to predict SLOR is particularly promising. By analyzing longitudinal changes in radiomic features alongside pharmacokinetic and clinical data, AI can detect patterns associated with waning drug efficacy before clinical relapse occurs[37]. For example, increasing bowel wall heterogeneity, changes in creeping fat distribution, or declining IFX trough levels may signal imminent loss of response. AI models can integrate these signals into risk scores, enabling clinicians to intervene early. Such proactive adjustments might include dose escalation, addition of immunomodulators, or switching to a biologic with a different mechanism of action. AI models could help reduce complications, avoid hospitalizations, and improve long-term outcomes[38]. Beyond predicting treatment response, AI can also inform broader aspects of IBD management. Models integrating radiomic and clinical data may help stratify patients for clinical trials, ensuring that therapies are tested in populations most likely to benefit. AI can also support drug development by identifying imaging biomarkers that serve as surrogate endpoints. Moreover, AI-driven decision-support tools could be embedded into clinical workflows, providing real-time guidance for therapeutic adjustments based on patient-specific risk profiles (Figure 1). Integration of radiomics and AI represents a powerful step toward precision medicine in IBD. By moving beyond reactive monitoring toward proactive, personalized care, these technologies offer the potential to improve outcomes, optimize resource use, and enhance patient quality of life.

Figure 1 Workflow illustrating the distinction between radiomic feature extraction and artificial intelligence/machine learning model development in inflammatory bowel disease management. The diagram highlights two sequential stages. The radiomics workflow (blue cluster) begins with image acquisition and preprocessing (e.g., computed tomography enterography, magnetic resonance enterography), followed by segmentation of regions of interest, calculation of quantitative features (such as shape, intensity, and texture), and feature selection or reduction. This process yields structured quantitative imaging features, which serve as standardized data inputs. The AI/machine learning workflow (orange cluster) uses these features – often combined with clinical and laboratory data – for model training, validation, and predictive modeling. Outputs from the artificial intelligence/machine learning stage include individualized predictions, such as risk of secondary loss of response or likelihood of treatment response. Finally, these outputs feed into clinical integration (green cluster), where predictions inform decision support, risk stratification, and optimization of therapeutic strategies in Crohn’s disease and ulcerative colitis. AI: Artificial intelligence; ML: Machine learning; SLOR: Secondary loss of response; IBD: Inflammatory bowel disease.

CHALLENGES IN IMPLEMENTATION

Despite the remarkable promise of radiomics and AI for improving the management of IBD, the path to clinical adoption is fraught with challenges. These obstacles span biological variability, technical limitations, data governance and ethics, and issues of clinical acceptance. Unless systematically addressed, they risk limiting the translation of these innovations from research into practice. A first and fundamental challenge arises from the heterogeneity of IBD itself. CD and UC are not singular entities but rather encompass a spectrum of phenotypes, behaviors, and disease locations. Patients with CD may present with inflammatory, stricturing, or penetrating disease, each associated with distinct imaging characteristics. Similarly, UC varies in extent, ranging from proctitis to pancolitis, with different risks of relapse and therapeutic responses. This inherent biological diversity complicates the development of radiomic and AI models that are robust across populations. A model trained predominantly on ileal CD may underperform when applied to colonic CD or to UC. Moreover, variability in treatment history, prior surgery, and concomitant therapies can introduce further complexity. To overcome these challenges, large and diverse training datasets are required, ideally spanning multiple centers and geographic regions. Collaborative consortia and data-sharing initiatives will be essential to capture the full breadth of disease heterogeneity.

Technical variability compounds the problem. Radiomic feature extraction is sensitive to differences in image acquisition, reconstruction, and preprocessing. In CTE, variations in slice thickness, contrast timing, or reconstruction algorithms can alter texture features, while in MRE, differences in field strength, sequence parameters, and coil technology influence signal characteristics[39,40]. Without harmonization, models trained on one set of parameters may fail to generalize. Initiatives such as the Image Biomarker Standardization Initiative have provided consensus definitions and benchmarks, but consistent adoption across centers remains limited. Preprocessing strategies such as image resampling and intensity normalization can mitigate some variability, yet reproducibility studies show that certain features remain unstable. A practical path forward requires not only adherence to consensus standards, but also systematic reporting of acquisition parameters and feature reproducibility metrics. Only with transparency and reproducibility can the field move toward regulatory acceptance and clinical confidence.

Another layer of complexity involves segmentation. Radiomic analysis depends on accurate delineation of regions of interest, such as bowel wall or mesenteric fat. Manual segmentation, while widely used, is time-consuming and subject to interobserver variability. Semi-automated approaches improve efficiency but may struggle in cases of poor contrast or complex anatomy. DL-based segmentation algorithms, such as U-Net variants, offer a promising solution, enabling rapid, reproducible delineation[41]. However, these models also require large annotated datasets for training and remain vulnerable to errors when confronted with atypical anatomy or artifacts. Robust evaluation of segmentation reproducibility, including intra- and interobserver studies, is essential for clinical translation.

Beyond technical reproducibility, there are challenges in validation and generalizability. Many published radiomics–AI models in IBD remain in the early stages of development, often relying on retrospective single-center datasets. Internal validation, while valuable, does not guarantee external performance. Temporal validation, which tests models on future patient cohorts, and external validation across independent centers, are critical for demonstrating generalizability. Without such validation, models risk being brittle, performing well in controlled research environments but poorly in real-world clinical practice. Prospective validation studies, ideally embedded in clinical workflows, are necessary to build confidence in clinical utility.

Data governance and ethical considerations present another formidable barrier. Radiomics and AI depend on access to large volumes of patient data, including imaging, clinical, and molecular information. This raises questions about patient privacy, data security, and informed consent. Regulations such as the General Data Protection Regulation in Europe and the Health Insurance Portability and Accountability Act in the USA impose strict requirements on data handling, storage, and sharing[42]. Anonymization and encryption are essential safeguards, but even deidentified imaging data may carry reidentification risks. Federated learning has emerged as one potential solution, enabling models to be trained across multiple centers without transferring raw data. However, federated approaches introduce new challenges, including harmonization of preprocessing pipelines and communication of model updates.

Ethical issues extend beyond privacy to questions of fairness and accountability. AI models trained on nonrepresentative populations risk embedding biases that could exacerbate health disparities. For example, a model trained predominantly on European populations may underperform in Asian or African cohorts, leading to inequities in care. Fairness assessments and subgroup performance analyses are essential, as is deliberate inclusion of diverse populations in training datasets. Transparency and interpretability are equally critical. Clinicians may be reluctant to trust “black box” models that produce predictions without explanation. Methods such as Shapley additive explanations or saliency maps can highlight which features contributed most to a prediction, supporting interpretability[43,44]. However, the challenge remains to balance accuracy with interpretability in ways that satisfy both clinicians and regulators.

Clinician adoption itself is a central challenge. Even the most accurate AI models have little impact if they are not used in practice. Adoption depends on trust, usability, and integration. Models must be transparent and interpretable, providing risk scores and actionable guidance aligned with clinical workflows. Outputs must be presented in a user-friendly manner, ideally integrated into electronic health records and radiology or gastroenterology dashboards. Clinicians must also be confident that responsibility for patient care remains clear: AI can support decision-making, but accountability ultimately lies with the clinician. Education and training are needed to familiarize clinicians with AI principles, limitations, and applications. Professional societies such as the European Crohn’s and Colitis Organization and the European Society of Gastrointestinal and Abdominal Radiology could accelerate adoption by issuing consensus statements, providing training modules, and integrating AI literacy into curricula.

Finally, regulatory approval and health-economic considerations cannot be overlooked. Regulators require evidence of safety, efficacy, and reproducibility, often through prospective trials. Health systems demand evidence of cost-effectiveness before adoption. Demonstrating that radiomics–AI models improve outcomes, reduce hospitalizations, or lower costs compared to standard care will be critical. Without such evidence, even the most promising models may struggle to cross the translational gap. All these challenges illustrate the complexity of integrating radiomics and AI into IBD care but they are not insurmountable. Through international collaboration, adherence to methodological standards, transparent reporting, and prospective validation, the field can move from proof-of-concept studies to clinically impactful tools. Addressing these challenges head-on is essential to realize the promise of precision medicine in IBD.

FUTURE DIRECTIONS

Looking ahead, the potential of radiomics and AI in IBD is vast, extending beyond current applications in treatment response prediction to encompass broader aspects of disease monitoring, stratification, and personalized care. Several key directions stand out as priorities for research and development. One of the most promising avenues is multimodal integration. While radiomics provides powerful insights into tissue heterogeneity and disease activity, it represents only one dimension of the disease. Genomic, transcriptomic, proteomic, metabolomic, and microbiome data provide complementary perspectives, capturing molecular and microbial drivers of inflammation[45]. Integrating multiomics layers with radiomic and clinical data, AI models could generate holistic disease profiles that better reflect the complexity of IBD. Such integrative models might predict treatment response and SLOR as well as risk of long-term complications, susceptibility to adverse events, and likelihood of extraintestinal manifestations[46]. The convergence of radiomics and multiomics could therefore usher in a new era of personalized IBD care, where therapy is tailored to the unique biological signature of each patient[35].

Real-time monitoring represents another exciting frontier. Traditional assessments of IBD rely on periodic clinic visits, laboratory tests, and imaging studies, which provide only snapshots of disease activity. Between visits, patients may experience fluctuations that go undetected until symptoms worsen. Digital health technologies, including wearable devices, smartphone applications, and home biomarker testing, could fill this gap by providing continuous streams of patient-generated data[47,48]. Wearables might capture physical activity, sleep, and heart rate variability, while apps could track symptoms, diet, and medication adherence. Home tests could monitor fecal calprotectin or inflammatory biomarkers. When integrated with radiomic and clinical data, these digital streams could feed into AI models that provide dynamic, real-time predictions of disease activity[49]. Clinicians could receive alerts of impending relapse, enabling pre-emptive adjustments. Patients could be empowered with personalized feedback, supporting self-management and engagement.

Clinical trials are an essential next step. To date, most radiomics–AI studies in IBD remain retrospective and exploratory. Prospective, multicenter trials are needed to test these models in real-world clinical settings. Such trials should evaluate predictive accuracy and clinical utility, whether the use of AI models improves outcomes compared to standard care. Randomized controlled designs may compare AI-supported treat-to-target strategies against conventional monitoring, measuring endpoints such as steroid-free remission, hospitalization rates, surgery, quality of life, and cost-effectiveness[50]. Adaptive trial designs may allow iterative refinement of models during the trial, accelerating learning and translation. Beyond traditional efficacy endpoints, these studies should also assess patient-reported outcomes and satisfaction, as adoption will depend on clinician acceptance as well as patient trust and engagement. Health-economic analyses will be critical. The promise of radiomics and AI extends beyond clinical outcomes to potential cost savings. Preventing SLOR and reducing hospitalization, surgery, and repeated endoscopic procedures could yield significant savings for healthcare systems. Cost-effectiveness models should evaluate not only direct healthcare costs but also indirect costs such as productivity loss and caregiver burden. Demonstrating value for money will be crucial for adoption by payers and health systems.

Another future direction is regulatory and methodological standardization. As AI models become more sophisticated, there is a pressing need for robust reporting and validation frameworks. Guidelines such as transparent reporting of a multivariable prediction model for individual prognosis or diagnosis-AI for reporting prediction models, Prediction Model Risk of Bias Assessment Tool for assessing risk of bias, and Checklist for Artificial Intelligence in Medical Imaging for imaging AI studies provide important benchmarks[51-53]. Future research should adhere to these frameworks, reporting feature reproducibility, validation strategies, calibration metrics, and decision-curve analyses[54]. Transparent code sharing, open datasets, and reproducible workflows will accelerate progress and build trust.

Global equity is also a vital consideration. IBD is a global disease, with incidence rising rapidly in Asia, South America, and Africa. However, most radiomics and AI studies to date have been conducted in high-income countries. Ensuring that models are trained on diverse populations and accessible to low- and middle-income countries is essential to avoid exacerbating global health disparities. Federated learning and cloud-based decision-support tools may offer pathways to broader accessibility, allowing models trained in one context to be adapted for others without requiring large local datasets.

Finally, interdisciplinary collaboration will be the engine driving future progress. Radiologists, gastroenterologists, surgeons, data scientists, bioinformaticians, ethicists, regulators, and patients must work together to co-design models that are clinically relevant, technically robust, ethically sound, and user friendly. Collaboration with industry will be important for scaling tools and embedding them into clinical workflows, but must be balanced with transparency and accountability. Academic–industry partnerships, supported by regulatory frameworks and patient advocacy, will be key to moving the field from pilot projects to widespread clinical impact. In summary, the future of radiomics and AI in IBD lies in multimodal integration, real-time monitoring, prospective trials, health-economic evaluation, global equity, and interdisciplinary collaboration. By pursuing these directions, the field can move beyond proof of concept to deliver tangible improvements in patient care.

CONCLUSION

The integration of radiomics and AI into IBD management represents an important advance toward precision medicine. By extracting quantitative imaging features that capture inflammation, fibrosis, and tissue heterogeneity, and combining them with clinical, laboratory, and pharmacokinetic data, radiomics and AI can generate predictive models that forecast treatment response and SLOR. These technologies hold the potential to shift care from reactive interventions to proactive, personalized strategies, guiding biologic optimization and reducing reliance on invasive monitoring. Their promise is evident in CD, where creeping fat and transmural changes can be quantified, and in UC, where mucosal inflammation and relapse risk can be assessed more objectively. However, significant challenges, including disease heterogeneity, technical variability, data governance, clinician adoption, and regulatory approval, must be overcome through rigorous validation, standardization, and interdisciplinary collaboration. As multimodal integration and real-time digital monitoring evolve, and as prospective trials demonstrate clinical utility and cost-effectiveness, radiomics and AI are poised to transform the long-term management of IBD, bringing the vision of truly personalized care closer to reality.