Deep Learning Is Rewriting How Doctors Diagnose and Treat Crohn’s and Colitis

Inflammatory bowel diseases affect millions of people worldwide, yet diagnosis still relies heavily on a clinician's subjective read of an endoscopy image or a pathologist's manual slide review, both prone to inter-observer disagreement. A comprehensive review published in June 2026 by Hughes, Bello, Liang, and colleagues maps how AI tools are now challenging that variability at virtually every stage of IBD care, from the first colonoscopy to multi-omics-guided treatment decisions. The evidence is promising, though most of it still comes from single-centre retrospective studies, so the field is moving fast without yet moving far enough.

Deep Learning Scores Mucosal Inflammation as Well as Expert Endoscopists

Endoscopic scoring is the cornerstone of IBD management, but validated scores like the Mayo Endoscopic Score and the SES-CD carry real inter-observer variability. AI models now automate this process with notable accuracy. A deep neural network called DNUC reached 90% accuracy identifying ulcerative colitis patients with a UCEIS score of zero, the remission threshold that increasingly drives treatment decisions. Recurrent neural networks trained on UC endoscopy videos matched human central readers with a kappa of 0.84, a level of agreement that would satisfy most clinical trial protocols.

Capsule endoscopy for Crohn's disease shows even sharper numbers. AI-assisted capsule endoscopy achieves an AUC of 0.99 for lesion detection in CD, with 94% sensitivity and 97% specificity. More practically, one system cut reading time from 96.6 minutes to 5.9 minutes for standard examiners, a reduction that matters enormously in a specialty where endoscopy backlogs are chronic.

Advanced imaging techniques extend the gains. A virtual chromoendoscopy-based model called the PICaSSO score correlated more strongly with histological indices than white-light endoscopy scores for UC mucosal healing. Confocal laser endomicroscopy systems distinguished Crohn's from UC with 97% sensitivity and 91% specificity, and an AI system for colitis-associated dysplasia achieved 90.4% lesion detection in prospective cohorts, a clinically meaningful target given the elevated colorectal cancer risk in long-standing IBD.

Automated Histology Could Standardize the Most Subjective Step in IBD Assessment

Histological remission is now considered the deepest and most durable treatment target in UC, yet reaching agreement on whether a biopsy qualifies is notoriously difficult. A meta-analysis cited in the review found AI-assisted histology models in UC reached 84% sensitivity and 87% specificity across studies, though expert pathologists still outperformed them on sensitivity alone. The gap is narrowing.

One multimodal system, an Endo-Histo fusion model that combines endoscopy video with histological images, reached 93% accuracy predicting histological remission at week 52, a clinically actionable timepoint for deciding whether to escalate or de-escalate therapy. Separate deep learning systems predicted postoperative recurrence in CD by identifying features in the muscular layer and myenteric plexus that human reviewers routinely overlook. AI applications in IBD histology are therefore moving beyond simple scoring toward predicting what happens after the scope comes out.

Cross-Sectional Imaging and Wearables Add Transmural and Real-World Layers

Colonoscopy and biopsies capture luminal disease; CT enterography, MRI enterography, and intestinal ultrasound reveal what is happening in the bowel wall and beyond. Radiomics models applied to CTE images now characterise CD activity and fibrosis and predict treatment response to both ustekinumab and infliximab. MRI-based AI enhances fibrosis detection and flags patients at risk of losing response to biologics before clinical relapse. An AI-driven intestinal ultrasound model detects bowel wall thickening with 90.1% accuracy and differentiates inflammatory from fibrotic strictures with 83.3% accuracy, a distinction that changes management from anti-inflammatory to surgical or endoscopic therapy.

Beyond the hospital, wearable devices monitoring physiological parameters and biomarkers including CRP, IL-6, and faecal calprotectin have predicted IBD flares up to seven weeks in advance in early studies. Remote monitoring platforms and e-health tools support continuity of care between clinic visits, particularly for patients in regions with limited specialist access. Natural language processing models applied to electronic medical records classify CD phenotype and treatment response with 85% accuracy and identify extraintestinal manifestations, such as joint and skin disease, with 94.1% accuracy, pulling signals from unstructured clinical notes that previously went unquantified.

Key Limitations: Single-Centre Cohorts and the Black-Box Problem

The review is direct about what the field still lacks. Most AI models have been validated in single-centre, retrospective cohorts, which limits how far their performance numbers generalise. Annotation practices, grading systems, and imaging methodologies vary so widely across studies that pooling results is difficult. Building the high-quality annotated datasets that train reliable models is expensive and slow.

The opacity of deep learning systems compounds the problem. Clinicians are understandably reluctant to act on a recommendation they cannot interrogate, and existing medical device regulations were not designed with AI-driven clinical tools in mind. Challenges for AI integration in IBD include data heterogeneity, regulatory gaps, cybersecurity concerns, and the need for explainability frameworks that translate model outputs into language a gastroenterologist can weigh against clinical context. Ethical concerns in AI for IBD add another layer: biases baked into training data can compromise algorithm fairness across demographic groups, and patient trust requires user-friendly interfaces, not just accurate outputs. Multicenter platforms like IBD Plexus and UR-CARE are beginning to address the standardisation gap, but the field is not there yet.

Multi-Omics Integration and AI-Designed Drugs Point Toward Precision IBD Care

The most forward-looking section of the review addresses what happens when AI stops analysing images and starts integrating biology. Machine learning models applied to genomics, transcriptomics, proteomics, metabolomics, and metagenomics data are beginning to stratify patients by molecular subtype, predict therapy response before treatment starts, and identify new drug targets. Microbiome intelligence analysis can phenotype pro-carcinogenic features and predict treatment response in ways that standard clinical variables cannot.

One concrete example from the pipeline: ISM012-042, a gut-restricted PHD inhibitor designed with AI assistance, showed promise in preclinical models and has entered phase I trials targeting intestinal barrier restoration. AI's role in advancing precision medicine for IBD is still emerging, but the convergence of multimodal imaging, molecular data, and large language models, including a dedicated tool called ChatIBD that delivers guideline-based management recommendations with greater than 95% accuracy, sketches a plausible path toward treating each patient's disease as the individual molecular entity it actually is, rather than a diagnostic category that two clinicians might score differently on the same day.

Evidence rating

EMERGING — 80/100

Caveats:

• This is a narrative review synthesizing many individual studies, not a single trial — the reported accuracy figures come from diverse studies with varying designs and populations.
• Most underlying AI models were validated in single-center, retrospective cohorts, which limits how broadly these results apply.
• Funding for this review was not disclosed by the authors.

Comprehensive narrative review synthesizing AI applications across IBD endoscopy, histology, cross-sectional imaging, and multi-omics data, drawing on literature through January 2026. Reported diagnostic performance metrics derive from heterogeneous individual studies, the majority of which are single-center and retrospective, limiting generalizability. No funding source was declared by the authors.