Integrated structural and functional brain imaging reveals biomarkers of disease activity in Crohn’s disease

Abstract

Background

The brain’s response to intestinal inflammation in Crohn’s disease (CD) remains poorly characterized. We aimed to identify a neuroimaging signature of disease activity by integrating structural and functional brain metrics with machine learning.

Methods

We analyzed data from 235 participants across two cohorts: a primary cohort of 180 individuals (75 healthy controls [HCs], 70 CD in remission [CDR], 35 active CD [CDA]) for model development, and an independent validation cohort of 55 CD patients (16 CDA, 39 CDR). All participants underwent 3D T1-weighted and resting-state functional MRI. We performed voxel-based morphometry and seed-based functional connectivity (FC) analyses. A machine-learning pipeline combining Lasso regression and support vector machine was used to select features and construct a classifier to distinguish CDA from CDR.

Results

Patients with CD exhibited widespread brain alterations compared to HCs. Critically, the gray matter volume (GMV) of the right inferior frontal gyrus (opercular part extending to the triangular part; rIFGoper) was significantly higher in CDA than in CDR and was positively correlated with fecal calprotectin levels. FC of the rIFGoper and bilateral putamen with the default mode and sensorimotor networks was diminished in CD and correlated with clinical disease activity. A classifier built on 10 key imaging features (including rIFGoper GMV) differentiated CDA from CDR with an area under the curve (AUC) of 0.85 in the internal test set and 0.73 in an independent external set.

Conclusions

We identified a distinct neuroimaging signature of active CD, characterized by structural enlargement of the rIFGoper alongside characteristic network dysconnectivity. This signature, which correlates with gut inflammation, demonstrates strong potential as an objective biomarker for stratifying disease activity.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12876-025-04547-x.

Introduction

Crohn’s disease (CD), a subtype of inflammatory bowel disease (IBD), is characterized by a relapsing-remitting course driven by chronic intestinal inflammation. While its incidence has plateaued in developed countries, it continues to rise in newly industrialized regions [1]. CD manifests with gastrointestinal symptoms such as abdominal pain, diarrhea, hematochezia, and malabsorption, which may involve extraintestinal manifestations, like arthritis. Clinical management is complicated by alternating phases of active inflammation (CDA) and remission (CDR), with therapeutic goals increasingly emphasizing sustained deep remission [2].

Emerging evidence implicates the brain–gut axis (BGA) in the pathogenesis and symptom expression of IBD [3–5]. Studies suggest that brain structural and functional alterations are not merely secondary to chronic pain or stress but may actively contribute to dysregulated intestinal immune response [6–10]. The central nervous system (CNS) may receive inflammatory input via the BGA, leading to plastic changes in the brain’s regulatory circuitry that, in turn, modulate gastrointestinal function in a bidirectional loop [11–14].

Structural MRI studies have shown gray matter volume (GMV) reductions in the frontal cortex, cingulate gyrus, and sensorimotor regions of CD patients compared to healthy controls (HCs) [3–5, 15]. Our previous study using resting-state fMRI focused on abdominal pain in CD patients and identified altered regional brain homogeneity in pain-related areas [16]. Similarly, a meta-analysis revealed reduced functional connectivity (FC) in the lower central lobule and cingulate gyrus in CD [17]. However, most of these studies enrolled patients in clinical remission, and comparative analyses across disease activity stages (CDA vs. CDR) remain limited. For instance, while Agostini et al. [18] reported decreased GMV and FC in patients with active CD, other studies failed to find significant group differences [15], highlighting the need for more systematic, multi-model investigations.

Accurate assessment of disease activity is critical for guiding CD management. Although the Crohn’s disease activity index (CDAI) is widely used, it remains subjective. Objective methods such as endoscopy or intestinal MRI are more accurate but less practical for frequent monitoring due to invasiveness or cost. Given the documented CNS involvement in CD, brain imaging may offer complementary, noninvasive biomarkers of disease activity that reflect both inflammatory and neuropsychological states. Yet, the potential of brain imaging as a surrogate marker of disease activity in CD is underexplored.

In parallel, machine learning approaches have demonstrated significant potential in biomedical imaging by uncovering complex and nonlinear patterns that may not be readily captured through conventional statistical techniques [19]. Prior studies have demonstrated the feasibility of using machine learning to classify disease states or predict prognosis based on multimodal neuroimaging and clinical features [20–22].

Therefore, we combined structural and resting-state functional MRI to examine brain alterations in CD patients across different disease activity stages. Specifically, we aimed to: (1) compare GMV and FC between CDA and CDR patients; (2) identify brain signatures associated with disease activity using both neuroimaging and clinical indices (e.g., CDAI, CRP); and (3) develop a machine learning-based classification model for distinguishing CDA from CDR based on integrated brain features. External validation was conducted to evaluate model generalizability.

Materials and methods

Study participants

This study was approved by the Institutional Review Board of Yueyang Hospital of Integrated Traditional Chinese and Western Medicine, Shanghai University of Traditional Chinese Medicine (Approval No. 2024 − 228), and all participants provided written informed consent. The study protocol was in compliance with the principles of the Declaration of Helsinki.

All patients were recruited from the IBD Clinic of Shanghai Institute of Acupuncture and Meridian, Shanghai University of Traditional Chinese Medicine, the Endoscopy Center of Zhongshan Hospital, Fudan University and Department of Gastroenterology, Ruijin Hospital, Shanghai Jiaotong University School of Medicine. All patients underwent systemic and gastrointestinal examinations (i.e., colonoscopy and pathologic biopsies), and the findings were confirmed by an experienced gastroenterologist. Laboratory tests and colonoscopy were both performed 2 weeks and 1 month before MRI examination. CRP levels, erythrocyte sedimentation rate, platelet levels, and CD endoscopic index of severity (CDEIS) scores were quantified for all patients. The CDEIS was used in this study under a copyright license obtained from the Groupe d’Étude des Affections Inflammatoires Digestives (GETAID) [23]. Information on date of birth, sex, and duration of education was also collected. In addition, according to the Montreal classification, information on the following parameters was collected: age at diagnosis (A), site of disease (L), disease behavior (B), perianal disease (p), extraintestinal manifestations, history of previous intestinal surgery, biologic therapy, and possible maintenance therapy. Furthermore, the severity of depressive and anxiety symptoms was assessed using the Hospital Anxiety and Depression Scale (HADS), and the quality of life was assessed using the Inflammatory Bowel Disease Questionnaire (IBDQ). The possibility of any patient having a psychiatric disorder was ruled out by a psychiatrist using DSV-IV. A total of 235 individuals were included in this study and divided into Cohort 1 [180 individuals (75 HCs, 70 CDR patients and 35 CDA patients) from the Shanghai Mental Health Center] and Cohort 2 [55 patients (16 CDA patients and 39 CDR patients) from the Longhua Hospital Center affiliated with the Shanghai University of Traditional Chinese Medicine].

The inclusion criteria for CD patients were (1) age 18–65 years, (2) duration of education ≥ 6 years, (3) disease activity or remission lasting ≥ 6 months, and (4) right-handed patients. The inclusion criteria for healthy subjects were (1) age 18–65 years, (2) duration of education ≥ 6 years, (3) no history of major illness or chronic disease, (4) no family history of a genetic disease, (5) no history of use of stimulant drugs in the past three months, and (6) right-handed individuals.

The exclusion criteria for CD patients were (1) history of CD-related abdominal surgery, (2) menstruating, pregnant, or lactating women, (3) history of use of corticosteroids, biological agents, or psychotropic drugs in the past 3 months, (4) claustrophobia, (5) presence of metal in the body or other contraindications to MRI examination, (6) history of neurosurgery (e.g., traumatic brain injury, cerebrovascular injury, head trauma, loss of consciousness, or other related diseases of nervous system dysfunction), and (7) family history of psychiatric or nervous system genetic diseases. The exclusion criteria for healthy subjects were menstruating, pregnant, or lactating women.

Disease grouping

The included CDAI scores, CDEIS, and laboratory test findings at baseline assessment were used to classify patients as having CDA or CDR. CDA was defined as (1) CDAI scores ≥ 150; (2) CDEIS ≥ 3; and (3) laboratory findings of serum CRP > 10 mg/L or fecal calprotectin ≥ 250 µg/g. CDR was defined as (1) CDAI scores < 150; (2) CDEIS < 3; and (3) laboratory findings of serum CRP ≤ 5 mg/L or fecal calprotectin ≤ 50 µg/g. Notably, a diagnosis of CDA or CDR required all three of their respective conditions to be met at the same time.

MRI data analysis

Data acquisition and scanning parameters

Notably, fMRI of Cohort 1 was performed in the Department of Radiology at the Shanghai Mental Health Center using a 3.0-Tesla MR scanner (Siemens Medical, Erlangen, Germany) equipped with a 12-channel head coil. Functional images were acquired with a single-shot gradient-recalled echo planar imaging sequence [time of repetition (TR)/time of echo (TE): 2000 ms/30 ms, field of view: 240 × 240 mm, matrix size: 64 × 64, flip angle: 90°, in-plane resolution: 3.75 × 3.75 mm, and 32 sagittal slices]. A set of high-resolution T1-weighted structural images was also collected (TR/TE: 2.3 s/2.98 ms, field of view: 256 × 256 mm, matrix size: 256 × 256, flip angle: 9°, in-plane resolution: 1 × 1 mm, slice thickness: 1.0 mm with no gaps, and 176 slices), which was consistent with our previous study.16 Furthermore, MRI (Ingenia 3.0T, Philips, The Netherlands) scanning equipped with an 8-channel head coil of Cohort 2 was performed at Longhua Hospital, Shanghai University of Traditional Chinese Medicine with the same scanning parameters and image processing, with the exceptions of TR/TE = 8.0 ms/3.7 ms and the number of scanning layers of T1 being 170.

Voxel-based morphometry analysis

The structural data were processed using the Statistical Parametric Mapping (SPM) and CAT12 (A Computational Anatomy Toolbox for SPM, https://neuro-jena.github.io/cat/) based on MATLAB 2023b, including segmentation and normalization processes using the Diffeomorphic Anatomical Registration using Exponentiated Lie (DARTEL) algebra algorithm, as described in previous studies [3, 24]. The DARTEL algorithm was chosen for its high accuracy in inter-subject anatomical alignment, which is crucial for reliable group comparisons. In detail, a customized anatomical template was built directly from participants’ T1-weighted images for normalization. The selected GM and white matter (WM) were extracted from the native brain and normalized to the corresponding tissue template. After the removal of non-brain regions, the original and the simulated images were segmented into GM, WM, and cerebrospinal fluid (CSF), and intensity inhomogeneity correction was performed. The SPM segmentation used a mixture-model cluster analysis to identify voxel intensities matching specific tissue types. All the segmented GM and WM images were normalized to GM and WM maps, respectively. An 8-mm full width at half-maximum isotropic Gaussian kernel was used to smooth the normalized GM images. This smoothing step improves the signal-to-noise ratio and accounts for residual anatomical variability, thereby increasing the sensitivity of subsequent group-level statistical tests. Finally, the statistical analysis was performed using a two-sample t-test within SPM12 to compare the GM volumes between CD and HC groups. The survival results were analyzed using a threshold of P < 0.001and underwent cluster-level multiple comparison correction (false discovery rate, FDR; P < 0.05).

Resting-state fMRI preprocessing

The resting-state images were processed using the DPABI (Data Processing & Analysis for Brain Imaging toolkit, https://rfmri.org/DPABI), which involved slice timing correction, head motion correction, spatial registration, normalization, smoothing, covariate removal, and filtering. In addition, subject exclusion criteria based on head movement information were applied such that data with translations greater than 2 mm or rotations exceeding 2 degrees were excluded.

ROI-based functional connectivity analysis and functional connectivity network construction

The regions showing differences in GMV were defined as regions of interest (ROIs). After preprocessing, the mean fMRI signal time series was extracted from each ROI. This involves averaging the fMRI signal across all voxels within each ROI for each time point. The resulting time series represent the neural activity of each ROI over time. Then, the FC map between each ROI and all voxels was computed using Pearson’s correlation coefficient and smoothed by a 6-mm Gaussian kernel.

The functional connectivity network was constructed using the Graph Theoretical Network Analysis (GRETNA, https://www.nitrc.org/projects/gretna/) toolkit. Initially, the entire brain was segmented into 164 regions according to the AAL3 Atlas. This comprehensive parcellation provides whole-brain coverage and a standardized anatomical framework, enabling network-wide analysis even when the focus is on specific regional connections. Using the preprocessed imaging data, Pearson correlation coefficients were computed for the average time series between each pair of brain regions to determine the functional connection strength. The resulting 164 × 164 R-value matrix was subsequently transformed into a Z-value matrix using Fisher’s r-to-z transformation, thus completing the construction of the whole-brain FC network.

Statistical analysis

Statistical analysis of clinical data was performed using the SPSS 22.0 software (IBM SPSS Inc., Chicago, Illinois, USA). Continuous variables were compared using Student’s t-test and categorical variables using the χ2 test. The Kolmogorov–Smirnov test was used to check the normality of the quantitative data before analysis. Data with normal distribution were expressed as mean ± SD and then evaluated using independent two-sample t-tests. The Mann–Whitney U test was used to analyze non-normally distributed data, which have been reported as medians and interquartile ranges. The differences among the three groups in the voxel-based morphometry analysis and ROI-based FC analysis were identified using SPM12, and the full factor analysis of the variance model was selected, with age and sex also used as covariates. A P value of < 0.001 was considered statistically significant, and in all the above analyses, multiple comparisons were corrected for FDR (P < 0.05, cluster-level). Pearson’s correlation analysis was used to study the relationship between the mean GMV in each ROI and the clinical variables (CDAI scores, CRP levels, and fecal calprotectin concentrations) for CD patients with sex and age as covariates. A similar analysis was also performed to observe the relationship between ROI-based FC and the clinical variables from the two-sample t-test results for each group. Given the potential long-term impact of prior medication use, a sensitivity analysis was conducted using gender, age, and medication history as covariates to mitigate potential confounding effects.

Machine learning classification

Dataset

The dataset includes two cohorts: Cohort 1: For each of the 105 patients from Shanghai Mental Health Center, 6 GMV values were extracted on the basis of group differences, including bilateral putamen, bilateral cerebellum, left ventral lateral thalamic nucleus, and right inferior frontal gyrus opercular part extending to the triangular part (rIFGoper), and 178 FC values were extracted as well. Cohort 2: For 55 patients from the Longhua Hospital Center, the same GMV and FC values were extracted as in Cohort 1.

Cohort 1 was used for model development and internal validation, whereas Cohort 2 was used for external validation of the constructed classifier to assess the model’s generalizability.

Feature selection and classifier construction

Cohort 1 was randomly divided into a training set and an internal test set with a ratio of 7:3, preserving the original distribution of patients with CDA and CDR. Given that the class distribution (approximately 1:2 CDA to CDR) reflects the clinical reality and was not deemed severe, we proceeded without explicit class-balancing techniques (e.g., synthetic minority over-sampling technique [SMOTE] or weighted loss functions). The model’s performance on the held-out internal test set confirmed that no significant bias toward the majority class was present. All subsequent steps, including data preprocessing, feature selection, and model training, were performed exclusively on the training set to prevent data leakage.

To mitigate potential systematic biases from the use of different MRI scanners (Siemens and Philips) with varying TR/TE parameters, Z-score normalization was applied at both the image intensity and feature levels. Any missing values were handled by mean imputation, with the mean value calculated from the training data. Feature selection was then conducted exclusively on the training set using a combination of a two-sample t-test (P < 0.05) and the least absolute shrinkage and selection operator (Lasso) algorithm. The Lasso regularization parameter was optimized via five-fold cross-validation within the training set to determine the optimal feature subset, ensuring an unbiased selection process and preventing leakage from the validation folds.

A support vector machine (SVM) classifier was constructed using the selected features. To assess the stability of the feature selection process, we performed 50 repeated cross-validation runs within the training set. The integrated pipeline—from preprocessing and feature selection to classification—was evaluated on the untouched internal test set from Cohort 1, with performance quantified by the area under the receiver operating characteristic curve (AUC). The finalized model was then prospectively applied to the independent external test set (Cohort 2) to assess its generalization capability.

Results

Patient characteristics

The clinical and demographic characteristics of all subjects are shown in Table 1. There were no significant differences in demographic characteristics (i.e., sex, age, height, and weight) among CDA, CDR, and HCs. There were no significant differences in disease course and the Montreal classification between CDA and CDR. The HADS-anxiety score (HADS-A) and HADS-depression score (HADS-D) were significantly higher in CD patients than in HCs (both, P < 0.05); however, they did not significantly differ between CDA and CDR. The CRP level was significantly higher in CDA than in CDR (P < 0.05), and the score for the IBDQ was significantly lower in CDA than in CDR (P < 0.05). A history of medication use was present in 25 CDA and 51 CDR patients, with no significant difference between the groups. Among these, 9 and 22 patients were treated with 5-aminosalicylic acid, respectively, while 9 and 11 received azathioprine. Combination therapy with 5-aminosalicylic acid and azathioprine was used by 3 and 7 patients, respectively. Methotrexate was used by one patient in each group. Furthermore, 3 and 10 patients had been treated with biologic therapy.

Table 1.

Clinical and demographic characteristics of patients with active crohn’s disease, those in remission, and healthy controls

	CDA (n = 35)	CDR (n = 70)	HCs (n = 75)	P value
Sex (male/female)	25/10	54/16	46/29	0.114
Age (years)	34.91 ± 11.47	31.66 ± 8.52	32.69 ± 8.33	0.226
Height (cm)	168.77 ± 6.75	170.60 ± 6.65	169.48 ± 6.86	0.377
Weight (kg)	56.19 ± 10.82	57.77 ± 9.37	59.54 ± 6.65	0.148
BMI (kg/m2)	19.67 ± 3.23*	19.78 ± 2.55*	20.73 ± 1.98	0.035
Duration of education (years)	16.26 ± 3.96*	15.97 ± 3.04*	17.63 ± 3.00	0.006
Duration of CD (months)	44.86 ± 36.45	58.07 ± 48.51	-	0.158
The Montreal classification				0.723b
A1	1	4	-
A2	26	60	-	0.114a
A3	8	6	-
L1	13	24	-
L2	6	8	-	0.686a
L3	16	38	-
B1	22	39	-
B2	9	27	-	0.489a
B3	4	4	-
B1p	7	12	-	0.790b
B2p	2	5	-	1.000b
B3p	2	1	-	0.257b
Medication	25	51	-	0.838
5-ASA	9	22	-	0.832
AZA	9	11	-	0.327
MTX	1	1	-	1.000 b
5-ASA + AZA	3	7	-	1.000 b
Biologic therapy history	3	10	-	0.542b
HADS-A	6.40 ± 3.83*	5.84 ± 3.45*	3.25 ± 2.23	< 0.001
HADS-D	5.20 ± 4.12*	4.41 ± 3.32*	2.76 ± 2.00	< 0.001
CDAI scores	212.62 ± 40.34	87.22 ± 44.72	-	< 0.001
IBDQ	160.17 ± 28.22	174.66 ± 27.83	-	0.014
CRP (mg/L)	12.16 ± 3.95	3.88 ± 2.01	-	< 0.001

a, Likelihood ratio; b, Fisher's exact test; -, not suitable; *, P < 0.05 vs. HCs

5-ASA 5-aminosalicylicacid, A1 >17, A2 17–40, A3 <40, AZA Azathioprine, B1 non-stricturing non-penetrating, B2 stricturing, B3 penetrating, BMI body mass index, CD Crohn’s Disease, CDA active Crohn's disease, CDAI Crohn’s Disease Activity Index, CDEIS Crohn's disease endoscopic index of severity, CDR Crohn's disease in remission, CRP C-reactive protein, HADS-A The Hospital Anxiety and Depression Scale - anxiety dimension, HADS-D The Hospital Anxiety and Depression Scale - depression dimension, HCs healthy controls, IBDQ Inflammatory Bowel Disease Questionnaire, L1 ileal, L2 colonic, L3 ileocolic, MTX methotrexate

Differences in gray matter volume

Age and sex were used as covariates, and the main effect differences among the three groups were located in the bilateral putamen. The post-hoc analysis showed that the GMV of the bilateral putamen was higher in CDA and CDR than in HCs. The left ventrolateral nucleus of the thalamus (lTHALvl) and bilateral cerebellar GMV were lower in CDA than in HCs. The GMV of the rIFGoper was higher in CDA than in CDR (P < 0.05, FDR corrected; Fig. 1).

Fig. 1.

Differences in gray matter volume between the three groups. The main effect of the gray matter volume (GMV) was different among the three groups, and the brain regions were located in the bilateral putamen (A). The differential brain regions of CDA and CDR patients were located in the rIFGoper (B); the differential brain regions of CDA patients and HCs were located in the bilateral putamen, lTHALvl, and bilateral cerebellum (C); and the differential brain regions of CDR patients and HCs were located in the bilateral putamen (D). A scatterplot (E) of the extracted GMV estimates in brain regions with significant between-group differences compared among the three groups. *P < 0.05; **P < 0.01; ***P < 0.001. lTHALvl, left ventrolateral thalamus; rIFGoper, inferior frontal gyrus opercular part extending to the triangular part

Differences in functional connectivity

The FC results primarily highlight differences between CD patients and HCs. Main effect analysis revealed decreased FC among 4 ROIs [rIFGoper, left thalamus (lThalamus), left putamen (lPutamen), and right putamen (rPutamen)] and between brain regions within the default mode network (DMN) and sensorimotor network (SMN).

Further post-hoc tests showed that compared to HCs, CD patients showed reduced FC between the rIFGoper and DMN regions [left cuneus (lCuneus), left lingual gyrus (lLingual), right inferior parietal lobule (rIPL), and left middle temporal gyrus (lMTG)] (Fig. 2A); compared to HCs, CD patients showed decreased FC between the lThalamus and right hippocampus and between the lThalamus and right thalamic lateral geniculate nucleus (rThalLGN). Furthermore, compared to HCs, CDA patients showed reduced FC between the lThalamus and bilateral cerebellum (Fig. 2B) and between the lPutamen and SMN regions [left insula (lInsula), left supplementary motor area (lSMA), left cerebellum, and rPutamen] (Fig. 2C); furthermore, they also showed decreased FC between rPutamen and SMN regions [rInsula, lSMA, left midcingulate cortex (lMCC), lThalamus, left postcentral gyrus (lPoCG), and left precentral gyrus (lPreCG)] (Fig. 2D). In addition, CDA patients showed reduced FC between the rPutamen and lInsula compared to CDR (P < 0.05, FDR corrected; Fig. 2D).

Fig. 2.

Differences in brain functional connectivity among the three subject groups. Compared to HCs, CDA and CDR patients both exhibited reduced functional connectivity in multiple brain regions. Shown are the main effect and post-hoc group differences in functional connectivity with the whole brain based on the regions of interest: right inferior frontal gyrus opercular part (rIFGoper) (A), left thalamus (lThalamus) (B), left putamen (lPutamen) (C), and right putamen (rPutamen) (D). Horizontal bars indicate the mean values. ^*P < 0.05; ^**P < 0.01; ^***P < 0.001. lCerebellum, left cerebellum; lCuneus, left cuneus; lInsula, left insula; lLingual, left lingual gyrus; lMCC, left midcingulate cortex; lMTG, left middle temporal gyrus; lSMA, left supplementary motor area; lThalLGN, left thalamic lateral geniculate nucleus; lPreCG, left precentral gyrus; lPoCG, left postcentral gyrus; rIPL, right inferior parietal lobule

Correlation analysis

With age and sex as covariates, in CDA, the GMV of the rIFGoper was positively correlated with fecal calprotectin concentrations (r = 0.37, P = 0.03; Fig. 3A). The FC of the rIFGoper with the left hippocampus (lHippo; r = -0.36, P = 0.04; Fig. 3B), left orbitofrontal cortex anterior part (lOFCant; r = -0.55, P < 0.001; Fig. 3C), left inferior temporal gyrus (lITG; r = -0.40, P = 0.02; Fig. 3D), and right orbitofrontal cortex lateral part (rOFClat; r = -0.35, P = 0.02; Fig. 3E) was negatively correlated with CDAI scores, and the FC between the rIFGoper and the right orbitofrontal cortex medial part (rOFCmed) was negatively correlated with CRP levels (r = -0.43, P = 0.01; Fig. 3F).

Fig. 3.

Correlation between brain imaging parameters and clinical and fecal inflammatory parameters of the rIFGoper. The rIFGoper gray matter volume was positively correlated with fecal calprotectin concentration (A). The functional connectivity between the rIFGoper and lHippo (B), lOFClat (C), lITG (D), and rOFCant (E) was negatively correlated with CDAI scores. The functional connectivity between the rIFGoper and rOFCmed was inversely correlated with serum CRP levels (F). The abscissa is the normalized residuals of clinical indicators (CDAI scores, CRP levels, and fecal calprotectin concentrations) with sex and age as covariates. The ordinate is the normalized residuals of the gray matter volume or brain functional connectivity with sex and age as covariates. rOFCant, right orbitofrontal cortex anterior part; lOFClat, left orbitofrontal cortex lateral part; lITG, left inferior temporal gyrus; rOFCmed, right orbitofrontal cortex medial part; lHippo, left hippocampus

Sensitivity analysis

To evaluate the robustness of our findings, we performed a series of sensitivity analyses. First, to assess the influence of potential confounding factors, we incorporated gender, age, and medication history as covariates in the comparisons of GMV and FC across the three groups. After controlling for these variables, the main effect of group difference in GMV of the rIFGoper became newly significant, with CDA patients showing significantly higher GMV in this region compared to HC. However, the GMV difference in the left thalamus between these two groups was no longer observed. Furthermore, FC analysis based on the AAL3 atlas revealed that although the number of brain regions with significant between-group differences increased, the primary connectivity patterns identified in the original analysis were consistently reproduced (Supplementary Tables 1–7). Second, we re-evaluated the correlation analyses from Sect. Correlation analysis using multiple linear regression with FDR correction. The results confirmed that the FC between the rIFGoper and the IOFCant remained significantly negatively correlated with the CDAI score (t = 5.18, FDR-corrected P = 0.017; Supplementary Table 8).

In summary, the sensitivity analyses confirm the robustness of the main findings of this study, particularly the structural alterations in the rIFGoper and its functional connectivity patterns. Detailed results are provided in the Supplementary Materials.

Machine learning classification performance

A total of 10 features were selected for model construction, comprising GMV of the rIFGoper, left thalamus, and left cerebellum, along with FC of the bilateral putamen and the rIFGoper with nodes of the DMN and SMN (Fig. 4A; Table 2). The five features with the highest absolute weights in the LassoCV model were: GMV of the rIFGoper, GMV of the left cerebellum, FC between the right putamen and the left inferior frontal gyrus triangular part (lIFGtri), FC between the left putamen and the cerebellar Vermis_6, and FC between the right putamen and the right middle frontal gyrus 2 (rMFG2) (Fig. 4B). All of these top-contributing features demonstrated perfect stability (selection frequency = 1.000) across 50 repeated cross-validation runs (Supplementary Table 9).

Fig. 4.

A set of 10 neuroimaging features effectively distinguished between CDA and CDR patients in both internal and external validation. These features comprised GMV of the rIFGoper, left thalamus, and left cerebellum, along with FC of the bilateral putamen and the rIFGoper with the DMN and SMN. Brain region map of the 10 features selected by LassoCV (A). Coefficients of the 10 features selected by LassoCV, with the top five features having the highest absolute weight coefficients (shown in orange) (B). The ROC curve of the SVM classifier showing brain imaging features distinguishing between CDA and CDR (AUC = 0.85) (C). The ROC curve of the SVM classifier on the external independent test set (AUC = 0.73) (D). rIFGoper, inferior frontal gyrus opercular part extending to the triangular part; lTHAL_VL: left ventrolateral nucleus of the thalamus; lCerebellum, left cerebellum; rPutamen, right putamen; lIFGtri; left inferior frontal gyrus triangular part; rMFG2, right frontal middle gyrus 2; lPutamen, left putamen; rIFG, right inferior frontal gyrus; lPoCG, left postcentral gyrus

Table 2.

The centroid of 10 feature values corresponding to brain regions

The brain region with difference in GMV	Centroid			The brain region with differences in FC with rIFGopoer as ROI	Centroid
	X	Y	Z		X	Y	Z
Right inferior frontal gyrus opercular part extending to the triangular part	45	24	19.5	Right frontal middle gyrus 2	39	9	57
Left ventrolateral nucleus of the thalamus	-18	-16.5	7.5
Left cerebellum	-34.5	-76.5	-19.5

The brain region with differences in FC with right putamen as ROI	Centroid			The brain region with differences in FC with left putamen as ROI	Centroid
	X	Y	Z		X	Y	Z
Left postcentral gyrus	-24	-42	57	Cerebellar vermis segment 6	3	-72	-21
Cerebellar vermis segment 7	3	-72	-24	Right temporal inferior gyrus	42	-9	-36
Right frontal middle gyrus 2	42	51	18
Left inferior frontal gyrus triangular part	-48	27	24

The SVM classifier, constructed using this set of 10 features, demonstrated strong classification performance. It achieved an AUC of 0.85 on the internal test set (Fig. 4C) and 0.73 on the independent external test set (Fig. 4D). Among the features, the GMV of the rIFGoper contributed the most to the model’s discriminative power. These results indicate that the selected neuroimaging features effectively distinguish between patients with CDA and CDR.

Discussion

This study provides integrative structural and functional neuroimaging evidence linking brain alterations to intestinal inflammation in CD, characterized by structural enlargement of the rIFGoper and functional decoupling involving the DMN and SMN. This integrated alteration not only correlates with objective markers of gut inflammation but also demonstrates strong potential as a biomarker for differentiating disease activity states. Specifically, patients with CDA exhibited significantly increased GMV in the rIFGoper compared with those in CDR, and this structural change was positively correlated with fecal calprotectin levels—an objective marker of mucosal inflammation. These findings suggest that rIFGoper enlargement may reflect an inflammation–related neuroplastic adaptation, offering new mechanistic insight into the gut–brain axis in CD.

The notable GMV increase in the rIFGoper in CDA counters the more common narrative of GMV reduction observed in many chronic inflammatory and autoimmune conditions [25]. While the precise mechanisms remain to be fully elucidated, several neurobiological processes associated with acute or early-stage neuroinflammation offer plausible explanations. First, the GMV expansion might reflect reactive gliosis, characterized by the activation and hypertrophy of astrocytes and microglia. In the context of peripheral inflammation, circulating pro-inflammatory cytokines can compromise the blood-brain barrier and initiate a neuroinflammatory response; unlike the degenerative gliosis seen in later disease stages, initial glial activation involves cellular swelling and proliferation, potentially leading to a transient volume increase [26]. Second, dendritic arborization and synaptogenesis may contribute to this change. The rIFGoper is critically involved in interoceptive awareness, cognitive control, and emotional processing [27–29]. Persistent nociceptive and interoceptive signals from the inflamed gut could thus drive neuroplastic adaptations in this region, such as increased dendritic complexity and synaptic remodeling, as a compensatory mechanism for managing heightened visceral alertness [30]. Finally, low-grade vasogenic edema resulting from increased vascular permeability [31] could cause subtle extracellular fluid accumulation, which would be detectable as a localized GMV increase on T1-weighted MRI before any overt tissue loss occurs.

Collectively, we postulate that the GMV increase in active CD may represent an early and potentially reversible neuroinflammatory-neuroplastic response to acute peripheral inflammation. This contrasts with the GMV reduction typically seen in chronic neurological and psychiatric conditions, which stem from prolonged neurotoxicity, accelerated apoptosis, and synaptic pruning. Therefore, the trajectory of brain changes in CD might be biphasic: an initial, adaptive enlargement during active inflammation driven by glial activation and neuroplasticity, potentially transitioning to a degenerative phase with GMV decline if inflammation persists.

The robustness of our central finding—rIFGoper GMV increase in CDA—was further confirmed by sensitivity analyses controlling for potential confounders including gender, age, and medication history. The persistence of this structural alteration after accounting for these variables strengthens the case for its specific link to active intestinal inflammation, rather than being a mere epiphenomenon of chronic illness or medication exposure. Similarly, the core pattern of rIFGoper functional dysconnectivity remained stable in these controlled analyses.

Despite this potential adaptive early-stage expansion, the functional implications of rIFGoper alteration appear to be detrimental. The rIFGoper plays a central role in integrating visceral sensory inputs via the vagus nerve and modulating immune and pain responses through top-down regulation [32–34]. Previous evidence indicates that reduced vagal tone may exacerbate intestinal inflammation in IBD [35]. Our findings suggest that increased rIFGoper GMV, accompanied by its decreased connectivity within the DMN, may reflect impaired interoceptive sensitivity and delayed recognition of inflammatory signals. This dysfunction could compromise the monitoring of visceral status, delay symptom awareness, and thereby perpetuate inflammatory activity.

In line with this, we observed decreased FC between the rIFGoper and DMN regions in CDA patients compared to HCs. The DMN supports self-referential thinking, emotional regulation, and integration of internal and external signals [36, 37]. Disrupted coupling between these systems may limit the brain’s ability to process gut–derived inflammatory cues, blunting the awareness of disease reactivation. This mechanism could explain why some CDA patients fail to recognize early relapse symptoms (e.g., abdominal discomfort, fatigue), leading to delayed treatment. The decoupled rIFGoper-DMN circuit may thus represent a critical neural substrate for impaired interoception in active CD and a potential target for interventions aimed at restoring brain–gut communication.

Beyond the rIFGoper, CDA patients exhibited reduced GMV in the cerebellum and thalamus, as well as diminished connectivity involving the bilateral putamen and sensorimotor regions. These regions are key nodes in networks mediating affective processing, motor control, and visceral pain perception [38–40]. Notably, the insula—a central interoceptive hub—also exhibited functional alterations, further supporting disrupted gut–brain signaling as a core pathophysiological feature of CD. Together, these results indicate that CD-related inflammation affects a distributed brain network encompassing cognitive–affective and sensorimotor systems, rather than isolated regions.

Our results are also consistent with prior literature linking chronic intestinal inflammation to affective disturbances. CDA patients had higher anxiety and depression scores, aligning with previous findings.¹⁵ Altered interoceptive processing may underpin these affective changes by amplifying emotional reactivity and reducing cognitive control over bodily states [41–43]. These results underscore the importance of addressing psychological well-being as an integral part of comprehensive CD management.

To evaluate the diagnostic potential of these neuroimaging features, we developed a SVM classifier based on 10 key GMV and FC metrics. The model distinguished CDA from CDR with high accuracy (AUC = 0.85 in internal validation, 0.73 in an independent external test set). Three features contributed most strongly to the classifier’s performance: increased rIFGoper GMV, reduced left cerebellar GMV, and decreased FC between the right putamen and left IFGtri. These results demonstrate the feasibility of using brain-based markers to objectively stratify disease activity and complement conventional inflammatory indices. With further refinement, such neuroimaging signatures could aid in early relapse detection, personalized monitoring, and therapeutic evaluation.

Several limitations of this study should be acknowledged. First, the modest sample size and inherent class imbalance, while reflecting clinical reality, may impact the stability of the findings. Second, the cross-sectional design prevents causal inferences, and residual confounding from factors like psychological comorbidities or disease duration cannot be excluded. Third, scanner-related batch effects likely contributed to the performance drop in external validation. Although we applied feature-level Z-score normalization as a pragmatic first step to mitigate this bias, future multi-center studies would benefit from employing more advanced harmonization techniques (e.g., ComBat or deep learning-based methods) to more effectively eliminate scanner-induced variance and enhance model generalizability. Additionally, while key findings were robust, some exploratory correlations did not survive multiple comparisons correction, and clinical biomarkers were not integrated into the predictive model. Future research should develop combined models that incorporate neuroimaging features with peripheral inflammatory markers and clinical indices, which hold strong promise for further improving the accuracy of disease activity prediction and relapse forecasting. Longitudinal studies combining serial neuroimaging with immune profiling are also essential to establish causality and investigate the reversibility of these brain alterations.

In summary, our findings establish a reproducible neuroimaging signature of active CD, centered on rIFGoper enlargement and related network-level dysconnectivity. This signature not only advances our pathophysiological understanding of brain-gut interactions but also translates into a promising biomarker framework for objectively assessing disease activity and predicting clinical course.

Conclusions

In conclusion, this study demonstrates that active CD is associated with a distinct and reproducible neuroinflammatory-neuroplastic brain signature, defined by rIFGoper structural enlargement within a context of dysconnectivity affecting both the DMN and SMN. These alterations may disrupt interoceptive processing and emotional regulation, impairing symptom awareness and potentially contributing to disease exacerbation. Neuroimaging markers—particularly GMV in the rIFGoper, cerebellum, and thalamus, and FCs involving the bilateral putamen—show strong potential to distinguish active from remitted disease states. These brain-based indicators provide mechanistic insight into gut–brain interactions and may serve as objective biomarkers for disease monitoring and relapse prediction. In clinical practice, interventions designed to enhance interoceptive awareness, emotional resilience, and brain-gut regulatory function could complement conventional anti-inflammatory therapies to improve long-term outcomes in CD.

Abbreviations

FC

Functional connectivity

CD

Crohn’s disease

CDA

Active CD

CDR

CD in remission

HCs

Healthy controls

GMV

Gray matter volume

DMN

Default mode network

SMN

Sensorimotor network

IBD

Inflammatory bowel disease

BGA

Brain–gut axis

CNS

Central nervous system

CDAI

Crohn’s disease activity index

CDEIS

CD endoscopic index of severity

GETAID

Groupe d’Étude des Affections

HADS

Hospital Anxiety and Depression Scale

IBDQ

Inflammatory Bowel Disease Questionnaire

TR

Time of repetition

TE

Time of echo

WM

White matter

CSF

Cerebrospinal fluid

ROIs

Regions of interest

rIFGoper

Right inferior frontal gyrus opercular part extending to the triangular part

Lasso

Least absolute shrinkage and selection operator

SVM

Support vector machine

AUC

Area under the receiver operating characteristic curve

HADS-A

HADS-anxiety score

HADS-D

HADS-depression score

lTHALvl

Left ventrolateral nucleus of the thalamus

lThalamus

Left thalamus

lPutamen

Left putamen

rPutamen

Right putamen

lCuneus

Left cuneus

lLingual

Left lingual gyrus

rIPL

Right inferior parietal lobule

lMTG

Left middle temporal gyrus

rThalLGN

Right thalamic lateral geniculate nucleus

rInsula

Right insula

lInsula

Left insula

lSMA

Left supplementary motor area

lMCC

Left midcingulate cortex

lPoCG

Left postcentral gyrus

lPreCG

Left precentral gyrus

lHippo

left hippocampus

lOFCant

Left orbitofrontal cortex anterior part

lITG

Left inferior temporal gyrus

rOFCant

Right orbitofrontal cortex anterior part

rOFCmed

Right orbitofrontal cortex medial part

lIFGtri

Left inferior frontal gyrus triangular part

rMFG2

Right middle frontal gyrus 2

lOFClat

Left orbitofrontal cortex lateral part

Authors’ contributions

Chunhui Bao: Conceptualization, Data curation, Funding acquisition, Investigation, Methodology, Writing – original draft; Shuai Xu: Formal analysis, Methodology, Software, Writing – original draft; Luyi Wu: Data curation, Funding acquisition, Writing – review & editing; Xinyi Zhu: Data curation, Formal analysis, Writing – review & editing; Di Wang: Data curation, Formal analysis, Investigation; Zhenjie Zhang: Formal analysis, Methodology, Writing – review & editing; Changcheng Qu: Formal analysis, Methodology; Yin Shi: Data curation, Supervision; Xiaoqing Zeng: Data curation, Investigation; Jie Zhong: Data curation, Investigation; Jianye Zhang: Investigation, Resources; Song Wang: Investigation, Resources; Rencheng Zheng: Methodology; Xiaoming Jin: Writing – review & editing; He Wang: Methodology, Supervision, Writing – review & editing; Huirong Liu: Resources, Supervision, Writing – review & editing; Huangan Wu: Conceptualization, Funding acquisition, Supervision, Writing – review & editing. All authors have read and approved the final version of the manuscript.

Funding

This work was supported by 2024 Shanghai Oriental Talent Plan Youth Project, Excellent Youth Fund of National Natural Science Foundation of China (No. 82422074), Special Clinical Research Project in the Health Industry of Shanghai Municipal Health Commission (No. 202340036), the Shanghai Rising-Star Program (No.19QA1408100) and the National Key Basic Research Program of China (No. 2015CB554500).

Data availability

The data supporting the findings of this study are available within the article. All code of the machine learning algorithm was implemented in Python. Code can be acquired on reasonable request to the corresponding author (Huangan Wu). All requests will be evaluated based on institutional and departmental policies.

Declarations

Ethics approval and consent to participate

This study was approved by the Institutional Review Board of Yueyang Hospital of Integrated Traditional Chinese and Western Medicine, Shanghai University of Traditional Chinese Medicine (Approval No. 2024-228), and all participants provided written informed consent. The study protocol was in compliance with the principles of the Declaration of Helsinki.

Consent for publication

No applicable.

Competing interests

The authors declare no competing interests.