Baseline multimodal clinical-radiomics-pathomics model predicts mucosal healing in Crohn’s disease after infliximab therapy

Abstract

BACKGROUND

Mucosal healing (MH) represents a crucial therapeutic objective in Crohn’s disease (CD). Infliximab (IFX) is extensively employed to induce MH; however, reliable instruments for the early prediction of MH remain scarce. Radiomics and pathomics have the capacity to extract high dimensional features from images and pathological specimens. We postulated that a multimodal model integrating radiomics, pathomics, and clinical data could precisely forecast MH in CD patients undergoing IFX therapy.

AIM

To develop and validate a multimodal model for predicting MH in patients with CD following IFX therapy.

METHODS

This retrospective investigation recruited 269 patients with confirmed CD (ileal: 102, with 71 for training and 31 for testing; ileocolonic: 167, with 116 for training and 51 for testing). Clinical variables, whole-slide image-derived pathomics features (used to calculate a pathomics score), and spectral computed tomography enterography derived radiomics features (used to calculate a radiomics score) were extracted. We developed single-modality and combined models and evaluated the performance of 14 models using receiver operating characteristic curve analysis.

RESULTS

Among 269 patients with CD, 43.1% of those with ileal CD and 33.5% of those with ileocolonic CD achieved MH after receiving IFX. Baseline albumin and fecal calprotectin were identified as independent predictors (both P < 0.05). The integrated radiomics-pathomics model attained the highest area under the curve: 0.915 in the ileal test cohort and 0.922 in the ileocolonic test cohort, which was significantly superior to the clinical or single modality models (DeLong test, all P < 0.05). Calibration, decision curve analysis, and clinical impact curve verified favorable fitness and clinical applicability.

CONCLUSION

A baseline radiomics-pathomics approach can noninvasively predict MH with high accuracy. This strategy provides complementary pathophysiological information for stratifying treatment response.

Key Words: Crohn disease; Mucosal healing; Radiomics; Pathomics; Infliximab

Core Tip: In response to the clinical requirement for an accurate, noninvasive predictive instrument for mucosal healing (MH) in Crohn’s disease (CD), this research endeavors to develop and validate a multimodal integrated model. The combined radiomics-pathomics model forecasts MH in ileal/ileocolonic CD, with area under the curve values of 0.915/0.922, surpassing those of other models. It offers a noninvasive, precise means for predicting MH in CD patients, thereby promoting personalized clinical decision-making and enhancing patient prognoses.

INTRODUCTION

Crohn’s disease (CD) is a chronic inflammatory disorder that can affect any segment of the gastrointestinal tract and is characterized by relapsing and remitting symptoms[1]. Current therapeutic goals emphasize deep, sustained remission to halt disease progression and prevent surgical complications[2]. Achieving endoscopic mucosal healing (MH) has been shown to be superior to clinical remission alone for reducing the risks of relapse, hospitalization, and surgery[3,4]. Tumor necrosis factor (TNF) plays a central pathogenic role in CD. Infliximab (IFX), an anti-TNF monoclonal antibody, promotes MH[5], which is defined endoscopically as the absence of ulceration[6,7]. Previous studies have evaluated the associations of fecal calprotectin (FC) and C-reactive protein (CRP) levels with MH following IFX therapy[8-10]. Predicting MH before initiating IFX therapy in CD limited to the ileum vs CD involving the ileocolonic region reflects site-specific differences in pathological mechanisms, treatment responses, and prognosis, enabling more accurate assessment of therapeutic efficacy and more individualized management.

Previous studies have identified systemic inflammatory markers [e.g., FC, CRP, erythrocyte sedimentation rate (ESR), and albumin (Alb)] as predictors of MH in CD[11-13]. However, individual serologic biomarkers generally lack adequate sensitivity and specificity for reliable MH prediction; for example, FC has a sensitivity of 0.623 and a specificity of 0.817[8]. Although computed tomography enterography (CTE) can be used to assess MH[8], it primarily captures morphologic features. In contrast, radiomics extracts hundreds to thousands of high-throughput quantitative features from medical images[14,15]. Although radiomics has been used to predict baseline MH in CD[16], digital pathology can provide molecular-level information related to gene expression. Moreover, the recent integration of histopathology with machine learning has enabled lesion genotyping, risk stratification, and outcome prediction[17-19]. Pathomics has been shown to predict postoperative outcomes in gastric cancer, and integrated radiomics-pathomics models have demonstrated utility for predicting large-volume cervical lymph node metastasis in clinically N0 papillary thyroid carcinoma and pathological complete response to neoadjuvant chemotherapy in locally advanced rectal cancer[20-22].

Despite these advances, a reliable multimodal approach that integrates clinical, radiomics, and pathomics features to enable early, baseline prediction of MH after IFX therapy in CD remains an important unmet need. Developing an accurate, noninvasive multimodal model for MH could help optimize treatment strategies, improve treatment success rates, and reduce healthcare resource use, addressing an urgent challenge in clinical practice. This study constructed a multimodal model integrating clinical, radiomics, and pathomics features to predict IFX-induced MH in CD at baseline.

MATERIALS AND METHODS

Patients and study design

This retrospective study was approved by the Ethics Committee of the First Affiliated Hospital of Anhui Medical University, with a waiver of informed consent (approval No. PJ2024-11-68). The study was conducted in accordance with the Declaration of Helsinki and ensured patient confidentiality, anonymity, and protection of patient rights. We enrolled patients with CD treated at our institution between March 2017 and February 2025.

The inclusion criteria were as follows: (1) A confirmed diagnosis of CD according to European Crohn’s and Colitis Organization guidelines[23]; (2) Endoscopy, spectral CTE, and histopathologic assessment performed within 1 month before IFX initiation, with an interval of ≤ 3 days between endoscopic and CTE examinations; (3) No prior exposure to biologic therapy; (4) Disease localized to the ileal or ileocolonic segments; and (5) No active intra-abdominal infection (e.g., abscess) or significant/opportunistic infection within the 3 months before enrollment. The exclusion criteria were as follows: (1) Missing endoscopic evaluation at 6-9 months after IFX therapy; (2) Non-protocol IFX dosing regimens or treatment duration; (3) Inadequate bowel preparation or nondiagnostic CTE image quality; (4) Isolated jejunal or colonic involvement; and (5) A history of colonic or ileal resection (Supplementary Figure 1).

Based on anatomical location, enrolled patients with CD were stratified into ileal (n = 102) and ileocolonic (n = 167) subgroups. Each subgroup was then randomly split into training and test sets at a 7:3 ratio (ileal: 71 training and 31 testing; ileocolonic: 116 training and 51 testing).

Baseline clinical variables included age, sex, disease duration since diagnosis, CRP, ESR, white blood cell, Alb, FC, perianal involvement, smoking status, prior exposure to corticosteroids or immunomodulators, Montreal classification, and the Harvey-Bradshaw index (HBI).

IFX was administered intravenously at 5 mg/kg during induction (weeks 0, 2, and 6), followed by maintenance dosing every 8 weeks. All patients underwent spectral CTE, ileocolonoscopy, and histopathologic evaluation at baseline, and endoscopic reassessment was performed 6-9 months after IFX initiation. Two gastroenterologists independently performed the endoscopic evaluations and assessed MH. Bowel preparation consisted of 3-4 L of a compound polyethylene glycol-electrolyte solution administered the evening before endoscopy. MH was defined as a Simple Endoscopic Score for Crohn’s Disease (SES-CD) < 3 at the 6-9 months follow-up in bowel segments that showed ulceration at baseline[9,24,25].

Whole-slide images

Hematoxylin and eosin (HE)-stained slides were retrieved from the archival records of patients with CD. All slides underwent an initial review by a senior gastrointestinal pathologist with 15 years of experience in CD. Slides with suboptimal staining quality were recut from paraffin blocks to a thickness of 5 μm and restrained. Slides were digitized using a KF-pro-005 whole-slide scanner (KFBIO Technology for Health) at 40 × objective magnification (0.5 μm/pixel resolution).

Regions of interests (ROIs) were independently selected by the same senior gastrointestinal pathologist according to predefined criteria: Intestinal mucosa with the most severe inflammatory infiltration and prominent structural abnormalities, including noncaseating granulomas, crypt distortion, villous atrophy, transmural inflammation, lymphoid aggregates, or fibrosis with muscularis propria hyperplasia. Images were captured at 200 × magnification (1000 × 1000 pixels, TIFF format). One ROI was selected for patients with ileal CD, and two ROIs were selected for patients with ileocolonic CD. Features from multiple ROIs were averaged to represent overall pathological characteristics. All images were independently reviewed by a second pathologist with at least 5 years of experience in gastrointestinal pathology. Discrepancies were resolved by consensus, and both pathologists were blinded to the clinical data.

Spectral CTE and semi-automatic segmentation

All patients underwent standardized bowel preparation consisting of a 12-hour fast followed by oral administration of 500 mL isosmotic mannitol solution at 45 minutes, 30 minutes, and 15 minutes before scanning. To minimize motion artifacts, 20 mg racemic anisodamine hydrochloride was administered intravenously 10 minutes before the scan. Iodinated contrast medium (320 mg I/mL, 1.5 mL/kg) was injected intravenously at 3.0 mL/second. Imaging was performed on revolution CT or discovery CT 750 scanners (GE Healthcare) in dual-energy spectral mode, with acquisition parameters detailed in Table 1. Enteric-phase and venous-phase images were acquired 45 seconds and 70 seconds after injection, respectively.

Table 1 Spectral computed tomography enterography scanning parameters.

Manufacturer	General-electric	General-electric
Scanner model	Revolution CT	Discovery CT750
Sequence	Axial	Axial
Gantry rotation time (second)	0.3	0.3
Tube voltage (kV)	80-140	80-140
Tube current (mA)	150-300	150-300
Detector collimation (mm)	256 × 0.625	64 × 0.625
Matrix	512 × 512	512 × 512
Pitch	1.375:1	1.375:1
Slice thickness (mm)	5.0	5.0
Slice spacing (mm)	5.0	5.0

CT: Computed tomography.

The deep learning-based nnU-Net framework, which has shown strong performance in biomedical segmentation tasks, was applied to venous-phase monochromatic dual-energy CT images at 50 keV. Although prior studies used nnU-Net only to segment CTE images from patients with CD and active inflammation (SES-CD > 2), its established performance supported its use in this study[26,27]. Images were reviewed using a window width of 400 Hounsfield units (HU) and a window level of 40 HU. We performed full three-dimensional segmentation of CD lesions on CTE using nnU-Net and retained only segments corresponding to pathologically confirmed regions.

First, all CTE images were reconstructed and reviewed using dedicated medical image-processing software (e.g., ITK-SNAP). Using predefined anatomic landmarks (e.g., the superior mesenteric artery, ileocecal valve, and colonic flexures), the bowel was divided into standardized segments: Duodenum, jejunum, ileum, cecum, ascending colon, transverse colon, descending colon, sigmoid colon, and rectum.

Second, during endoscopic biopsy, the precise anatomic location of each specimen was documented, including the segment name; distance from the anal verge (for colonic segments) or pylorus (for small-bowel segments), and adjacent landmarks (e.g., the ileocecal valve for the terminal ileum).

Finally, two radiologists (with 5 years and 10 years of abdominal imaging experience, respectively) and the senior gastrointestinal pathologist jointly matched CTE-segmented lesions to the corresponding biopsy segments. Lesions on CTE were first assigned to bowel segments using predefined landmarks and then matched to biopsy sites by comparing biopsy location, lesion size, and morphologic features (e.g., stenosis, wall thickening, and enhancement patterns) described on CTE and endoscopy. Discrepancies were resolved through multidisciplinary discussion.

Subsequently, a radiologist manually corrected the results, which were then reviewed and verified by another radiologist to ensure segmentation accuracy met subsequent feature-extraction requirements. The nnU-Net code is publicly accessible at: https://github.com/MIC-DKFZ/nnUNet.

Pathomics and radiomics feature extraction

Pathomics feature extraction was performed using CellProfiler (v4.1.3; https://cellprofiler.org). Selected HE-stained TIFF images were imported, and the UnmixColors module was used to separate the HE channels into grayscale components. The ColorToGray module then converted the HE images to single-channel grayscale using the combine method. Next, the measure colocalization module quantified pixel-wise colocalization using six features: Correlation, slope, overlap coefficient, Manders’ coefficients, Costes’ automated threshold, and the rank-weighted colocalization coefficient. Three additional modules were used to extract complementary features: (1) Measure-granularity: Analyzed granular patterns in hematoxylin, eosin, and merged HE grayscale images; (2) Measure image quality: Assessed: Focus: Focus score, local focus score, and power log logslope; Intensity: Mean, median, median absolute deviation, SD, and lower/upper quartile intensity; Thresholding: Optimal threshold; and (3) Measure texture: Computed 8 Haralick features (correlation, inverse difference moment, sum average, sum entropy, entropy, difference entropy, infomeas1, and infomeas2). A total of 313 pathomics features were extracted from the ileal and ileocolonic regions, respectively.

Radiomic feature extraction was performed using PyRadiomics (version 3.0.1; https://pypi.org/project/pyradiomics/). All images were iso-tropically resampled to 1 mm × 1 mm × 1 mm voxel dimensions. Original and wavelet-filtered features were extracted from semi-automatically segmented intestinal regions corresponding to histopathologically confirmed areas. The original features comprised three categories: (1) First-order statistics; (2) Shape-based descriptors; and (3) Texture features derived from: Gray-level co-occurrence matrices, gray-level run-length matrices, gray-level size zone matrices, and gray-level dependence matrices. A total of 1168 radiomic features were extracted from the ileal and ileocolonic regions, respectively.

Construction of the clinical model

Clinical parameters from the ileal and ileocolonic subgroups were analyzed using the χ² test for categorical variables and the Mann-Whitney U test for continuous variables. Univariate logistic regression was used to identify potential risk factors (P < 0.1), which were subsequently entered into multivariate logistic regression models to determine independent predictors (P < 0.05). Based on these analyses, two distinct clinical prediction models were developed: Model 2 for ileal involvement and model B for ileocolonic involvement.

Feature selection and model development

The model development workflow is detailed in Figure 1. Both pathomics and radiomics features were standardized using z-scores. Feature selection and dimensionality reduction were performed independently for each modality using the training set in a three-step process: (1) Features with Pearson correlation coefficients ≥ 0.8 were excluded to minimize multicollinearity; (2) The select K best algorithm retained features with P ≤ 0.05; and (3) Least absolute shrinkage and selection operator (LASSO) regression identified optimal features with high relevance and low redundancy.

Figure 1 Combined models workflow implemented in this study. FC: Fecal calprotectin.

The pathomics score (pat-score) and radiomics score (rad-score) were computed for each patient as linear combinations of the selected features, each weighted by its LASSO coefficient. These scores were then incorporated into multivariable logistic regression models to develop: Pathomics-based models included model 3 (ileal) and model C (ileocolonic). Radiomics-based models included model 1 (ileal) and model A (ileocolonic).

Construction, evaluation, and validation of a nomogram

Multivariable logistic regression was performed to identify independent predictors among clinical risk factors, the pat-score, and/or the rad-score for nomogram development. For ileal involvement, four integrated models were constructed: Radiomics + clinical (model 4), pathomics + clinical (model 5), radiomics + pathomics (model 6), and combined (model 7). For ileocolonic involvement, four integrated models were constructed: Radiomics + clinical (model D), pathomics + clinical (model E), radiomics + pathomics (model F), and combined (model G).

Model discrimination was quantified using the area under the curve (AUC) of receiver operating characteristic (ROC) in both the training and testing sets. Calibration curves were used to assess goodness-of-fit across risk deciles. The DeLong test was used to compare AUCs between models. Decision curve analysis (DCA) evaluated clinical utility by plotting net benefit against threshold probabilities. When the threshold probability was 5%-30%, the model’s net benefit was substantially higher than the baselines of “no intervention (none)” or “treat all”, indicating clinical utility for decision-making. Clinical impact curves (CICs) visualized the relationship between predicted risk strata and observed outcomes.

Explanation of the model

The SHapley Additive exPlanations (SHAP) method, grounded in cooperative game theory and Shapley values, quantifies the direction and magnitude of each feature’s contribution to model predictions. This approach identifies key predictive features and clarifies the model’s decision logic. Through interactive visualizations, SHAP maps feature-specific contributions to individual predictions, thereby enhancing the interpretability of the model’s reasoning process. Consequently, SHAP improves model transparency and supports clinical trust.

Statistical analysis

Continuous variables were summarized as the mean ± SD or median (interquartile range), depending on the distribution. Between-group comparisons used the independent samples t-test for normally distributed continuous variables and the Mann-Whitney U test for non-normally distributed continuous variables. Categorical variables were analyzed using Pearson’s χ² test or Fisher’s exact test, as appropriate. All statistical analyses were performed using SPSS (version 27.0), R (version 4.3.3; https://www.Rproject.org), and Python (version 3.6.5). Statistical significance was defined as a two-tailed P < 0.05.

RESULTS

Clinical features and construction of the clinical model

Table 2 summarizes the baseline clinical characteristics of 102 patients with ileal CD and 167 with ileocolonic CD. MH was achieved by 44 (43.1%) patients in the ileal subgroup and 56 (33.5%) in the ileocolonic subgroup at 6-9 months after IFX. Univariate analysis identified variables associated with MH (P < 0.1): Sex, baseline Alb, perianal involvement, and HBI in ileal CD, and baseline CRP and FC in ileocolonic CD. Multivariate regression identified baseline Alb (ileal, P = 0.041) and FC (ileocolonic, P = 0.031) as independent predictors of MH (Table 3). These variables were used to develop the clinical prediction models: Model 2 (ileal) and model B (ileocolonic), with training/testing AUCs of 0.724/0.669 and 0.756/0.636, respectively (Table 4).

Table 2 Clinical factors in the training and testing sets, mean ± SD.

Clinical factors	Ileal (n = 102)			Ileocolonic (n = 167)
	Train (n = 71)	Test (n = 31)	P value	Train (n = 116)	Test (n = 51)	P value
Gender			0.387			0.989
Male	56	22		82	36
Female	15	9		34	15
Age at CD diagnosis (year), median IQR	28 (20-38)	30 (21-35)	0.951	24 (19-31)	27 (20-39)	0.126
Disease duration (day), median IQR	5 (2-23)	5 (1-58)	0.835	4 (2-16.75)	3 (1-15)	0.446
CRP at baseline (mg/L), median IQR	5.27 (2.17-17.06)	4.6 (1.30-19.91)	0.991	13.29 (3.19-36.30)	9.49 (4.25-25.20)	0.541
ESR at baseline (mm/hour), median IQR	16.25 (12.00-25.00)	22.00 (7.00-36.00)	0.555	30.40 (17.63-43.75)	16.00 (11.50-31.00)	< 0.001
WBC at baseline (× 109/L)	6.32 ± 2.09	6.31 ± 2.29	0.977	7.34 ± 2.23	6.46 ± 1.78	0.02
Alb at baseline (g/L)	39.66 ± 4.61	38.65 ± 4.73	0.313	37.68 ± 5.88	37.36 ± 4.79	0.74
FC at baseline (μg/g), median IQR	762.58 (394.78-1085.72)	774.94 (429.66-1137.76)	0.952	1086.68 (850.76-1335.94)	1019.20 (798.78-1263.00)	0.078
Perianal involvement			0.641			0.667
-	40	19		45	18
+	31	12		71	33
Smoking status at diagnosis			0.858			0.46
-	64	29		115	49
+	7	2		1	2
Previous steroids or immunomodulators use			0.858			0.478
0	46	21		82	32
1	21	9		26	16
2	4	1		8	3
Montreal behavior			0.355			0.811
Nonstricturing, nonpenetrating (B1)	47	20		98	41
Stricturing (B2)	7	6		11	6
Penetrating (B3)	17	5		7	4
HBI at baseline, median IQR	2 (1-4)	2 (1-5)	0.28	4 (1-6)	3 (1-5)	0.425
MH			0.871			0.971
MH	31	13		39	17
Non-MH	40	18		77	34

CD: Crohn’s disease; CRP: C-reactive protein; ESR: Erythrocyte sedimentation rate; IQR: Interquartile range; Alb: Albumin; FC: Fecal calprotectin; WBC: White blood cell; HBI: Harvey-Bradshaw index; MH: Mucosal healing.

Table 3 Positive results of univariate and multivariate logistic regression analysis for predicting mucosal healing in Crohn’s disease.

Features	Ileal						Ileocolonic
	Univariate			Multivariate			Univariate			Multivariate
	OR	95%CI	P value	OR	95%CI	P value	OR	95%CI	P value	OR	95%CI	P value
Gender	0.250	0.064-0.983	0.047	0.079	0.006-1.061	0.055	1.908	0.834-4.364	0.126
Age at CD diagnosis	1.029	0.989-1.071	0.155				1.009	0.968-1.053	0.662
Disease duration (day)	1.000	0.997-1.003	0.875				1.001	0.998-1.003	0.669
CRP at baseline	1.002	0.971-1.034	0.896				1.018	1.000-1.036	0.051	1.010	0.987-1.032	0.397
ESR at baseline	0.972	0.931-1.015	0.195				1.000	0.981-1.019	0.997
WBC at baseline	0.826	0.650-1.049	0.117				1.078	0.904-1.285	0.404
Alb at baseline	0.880	0.785-0.986	0.023	0.893	0.754-1.058	0.041	0.997	0.934-1.065	0.931
FC at baseline	0.999	0.998-1.000	0.142				0.999	0.998-1.000	0.015	0.999	0.998-1.001	0.031
Perianal involvement	0.443	0.170-1.159	0.095	0.247	0.051-1.187	0.081	1.352	0.617-2.965	0.451
Smoking status at diagnosis	1.037	0.214-5.016	0.964				828993669.358	0.000	1.000
Previous steroids or immunomodulators use	3.491	1.096-11.124	0.104				1.485	0.559-3.948	0.711
Montreal behavior	1.058	0.343-3.262	0.752				3.187	0.369-27.570	0.566
HBI at baseline	0.799	0.642-0.966	0.059	0.935	0.629-1.389	0.740	1.090	0.942-1.262	0.247
Rad score				1.173	1.010-1.363	0.037				3.820	1.541-9.473	0.004
Pat score				6.135	2.282-16.492	< 0.001				3.992	2.198-7.250	< 0.001

OR: Odds ratio; CI: Confidence interval; CD: Crohn’s disease; CRP: C-reactive protein; ESR: Erythrocyte sedimentation rate; Alb: Albumin; FC: Fecal calprotectin; WBC: White blood cell; HBI: Harvey-Bradshaw index.

Table 4 Results of all models predictive ability for predicting mucosal healing in Crohn’s disease.

Models		AUC (95%CI)	Accuracy	Sensitivity	Specificity	PPV	NPV
Model 1	Train	0.921 (0.860-0.982)	0.859	0.925	0.774	0.841	0.889
	Test	0.799 (0.641-0.957)	0.774	0.778	0.769	0.824	0.714
Model 2	Train	0.724 (0.616-0.833)	0.690	0.650	0.742	0.765	0.622
	Test	0.669 (0.486-0.851)	0.677	0.667	0.692	0.750	0.600
Model 3	Train	0.931 (0.874-0.989)	0.873	0.875	0.871	0.897	0.844
	Test	0.863 (0.703-1.000)	0.742	0.611	0.923	0.917	0.632
Model 4	Train	0.773 (0.656-0.889)	0.747	0.725	0.774	0.806	0.895
	Test	0.731 (0.533-0.928)	0.645	0.611	0.692	0.733	0.615
Model 5	Train	0.910 (0.845-0.975)	0.845	0.925	0.742	0.822	0.885
	Test	0.872 (0.744-1.000)	0.807	0.778	0.846	0.875	0.733
Model 6	Train	0.980 (0.955-1.000)	0.944	0.975	0.903	0.929	0.966
	Test	0.915 (0.808-1.000)	0.774	0.833	0.692	0.790	0.750
Model 7	Train	0.860 (0.778-0.942)	0.803	0.925	0.871	0.882	0.896
	Test	0.812 (0.668-0.956)	0.774	0.944	0.769	0.824	0.833
Model A	Train	0.873 (0.803-0.944)	0.819	0.805	0.846	0.912	0.659
	Test	0.827 (0.716-0.938)	0.706	0.618	0.882	0.913	0.647
Model B	Train	0.756 (0.674-0.839)	0.675	0.714	0.615	0.722	0.522
	Test	0.636 (0.518-0.754)	0.618	0.412	0.882	0.618	0.429
Model C	Train	0.951 (0.915-0.988)	0.922	0.961	0.846	0.925	0.917
	Test	0.866 (0.748-0.984)	0.804	0.882	0.647	0.833	0.733
Model D	Train	0.739 (0.636-0.842)	0.776	0.896	0.744	0.793	0.724
	Test	0.687 (0.501-0.873)	0.765	0.882	0.765	0.790	0.692
Model E	Train	0.919 (0.870-0.968)	0.836	0.792	0.923	0.953	0.784
	Test	0.819 (0.682-0.957)	0.720	0.697	0.765	0.852	0.667
Model F	Train	0.951 (0.916-0.987)	0.897	0.896	0.897	0.945	0.814
	Test	0.922 (0.843-1.000)	0.880	0.909	0.824	0.909	0.824
Model G	Train	0.950 (0.912-0.988)	0.871	0.844	0.923	0.956	0.750
	Test	0.908 (0.822-0.994)	0.820	0.788	0.882	0.929	0.682

Ileal: Model 1-radiomics, model 2-clinical, model 3-pathomics, model 4-radiomics and clinical, model 5-pathomics and clinical, model 6-radiomics and pathomics, model 7 combined. Ileocolonic: Model A-radiomics, model B-clinical, model C-pathomics, model D-radiomics and clinical, model E-pathomics and clinical, model F-radiomics and pathomics, model G-combined. CI: Confidence interval; AUC: Area under the curve; NPV: Negative predictive value; PPV: Positive predictive value.

Feature extraction, selection, pat-score, and rad-score establishment

For ileal CD, 313 pathomics features were extracted from pathological images. Sequential feature selection in the training set retained: (1) 64 features after Pearson correlation filtering (r < 0.8); (2) 19 features via select K best (P ≤ 0.05); and (3) 4 optimal features through LASSO regression. Similarly, 1168 radiomics features from CTE underwent identical selection: 192 features to 62 features to 16 features.

The resulting features generated ileal pat-score and rad-score (Figure 2). Pathomics model 3 and radiomics model 1 demonstrated training/testing AUCs of 0.931/0.863 and 0.921/0.799, respectively (Table 4). For ileocolonic CD, identical extraction yielded 313 pathomics features with training set selection: 122 features to 38 features to 13 features. And 1168 radiomics features: 216 features to 42 features to 7 features. These formed ileocolonic pat-score and rad-score (Supplementary Figure 2). The pathomics model (model C) and radiomics model (model A) achieved training/testing AUCs of 0.951/0.866 and 0.873/0.827, respectively (Table 4).

Figure 2 Feature selection and performance evaluation of the radiomics score (rad score) and pathomics score (pat score) in ileal Crohn’s disease via least absolute shrinkage and selection operator logistic regression. A: Selection of the tuning parameter (λ) in the least absolute shrinkage and selection operator model through 10-fold cross validation based on the minimum criterion. The binomial deviance from cross validation is plotted against the log transformed λ [ln(λ)]; B: Coefficient profiles of all features in relation to ln(λ). Features with non-zero coefficients subsequent to regularization are presented; C: The 20 features with non-zero coefficients and their respective contributions to the rad score and pat score; D: The verified sets of the rad score and pat score. LASSO: Least absolute shrinkage and selection operator.

Nomogram establishment, evaluation, and validation

Multivariable logistic regression identified baseline Alb, the pat-score, and the rad-score as independent predictors of MH in ileal CD, whereas FC, the pat-score, and the rad-score were independent predictors in ileocolonic CD (Table 3). These predictors were integrated to develop combined models for both regions (Figure 3A and B). As shown in Table 4, four models were constructed for ileal CD: Model 4 (training/testing AUC: 0.773/0.731), model 5 (0.910/0.872), model 6 (0.980/0.915), and model 7 (0.860/0.812). For ileocolonic CD, four models were constructed: Model D (0.739/0.687), model E (0.919/0.819), model F (0.951/0.922), and model G (0.950/0.908). Calibration curves demonstrated excellent fit for combined model 7 (ileal) and model G (ileocolonic) (Figure 3C and D). DCA (Figure 4A and B) showed net benefit for all models across threshold probabilities (0.1-0.9); model 6 and model F outperformed the others. CIC (Figure 4C and D) visualized decreasing high-risk patient counts and event frequencies with increasing risk thresholds, validating model 7 and model G’s clinical utility.

Figure 3 Radiomics and pathomics nomograms and calibration curves. A and B: Radiomics and pathomics nomograms for predicting Crohn’s disease mucosal healing status in the ileal and ileocolonic region; C and D: In the ileal and ileocolonic region, calibration curves of the combined models in the testing set. Alb: Albumin; FC: Fecal calprotectin.

Figure 4 Decision curve analysis and clinical impact curves. A and B: In the ileal and ileocolonic region, decision curve analysis of all models; C and D: Clinical impact curves of the combined models.

Explanation of the model

To enhance clinical interpretability, we computed Shapley values for population-level and patient-specific predictions in both combined models. Population-level feature contributions were visualized using beeswarm plots (Figure 5), which revealed differences in the impact of feature categories: Ileal: Pathomics 19% (4/21), radiomics 76% (16/21), and clinical 5% (1/21); Ileocolonic: Pathomics 62% (13/21), radiomics 33% (7/21), and clinical 5% (1/21). Feature values were color-mapped (red: High; blue: Low), and positive Shapley values (> 0) indicated non-MH-promoting effects. For individual predictions, Supplementary Figure 3 presents SHAP waterfall plots for four representative cases, delineating the direction and magnitude of each feature’s contribution. The baseline value [E f(x)] represents the cohort’s mean predicted probability, whereas f(x) denotes the individual patient’s predicted probability.

Figure 5 The SHapley Additive exPlanations beeswarm plot provides a comprehensive perspective on feature importance. Each row represents a single feature, and the horizontal axis indicates the corresponding SHapley Additive exPlanations (SHAP) value. Individual data points represent patient samples; red indicates higher feature values, while blue represents lower values. A positive SHAP value (> 0) is associated with a reduced likelihood of mucosal healing, whereas a negative value (< 0) is correlated with an elevated probability of mucosal healing. A: This plot illustrates 21 key features, including 4 pathomics features, 16 radiomics features, and 1 clinical feature, along with their combined impacts on the prediction probability. The three most influential features are: Granularity 14 Eosin, wavelet.LHL.glcm.ldmn, and diagnostics.lmage.interpolated.Maximum; B: This plot illustrates 21 key features, including 13 pathomics features, 7 radiomics features, and 1 clinical feature, along with their combined impacts on the prediction probability. The three most influential features are: Granularity 3 Eosin, log.sigma.0.5.mm.3D. firstorder. Mean, and X2Granularity 12 Eosin. They play important roles in the interpretability of the model and feature correlation. Alb: Albumin; FC: Fecal calprotectin; SHAP: SHapley Additive exPlanations.

Performance comparison between models

Figure 6 presents the ROC curves. Model 6 (ileal) and model F (ileocolonic) achieved the highest AUCs in both the training and testing sets, indicating superior performance for predicting MH in CD. As shown in Table 4, for ileal model 6, the training set metrics were accuracy 0.944, sensitivity 0.975, and specificity 0.903, and the testing set metrics were accuracy 0.774, sensitivity 0.833, and specificity 0.692. For ileocolonic model F, the training set metrics were accuracy 0.897, sensitivity 0.896, and specificity 0.897, and the testing set metrics were accuracy 0.880, sensitivity 0.909, and specificity 0.824.

Figure 6 In the ileal and ileocolonic region, performance of all models in the train set and testing set. A and B: Train set; C and D: Testing set. AUC: Area under the curve; CI: Confidence interval.

Performance comparisons consistently demonstrated the benefit of multimodal integration. In both regions, the radiomics + pathomics models (model 6: AUC = 0.915; model F: AUC = 0.922) outperformed the combined models (model 7: AUC = 0.812; model G: AUC = 0.908). In ileal CD, the pathomics + clinical model (model 5: AUC = 0.872) also outperformed model 7, whereas in ileocolonic CD, model G exceeded the pathomics + clinical model (model E: AUC = 0.819).

In ileal CD, model 5 outperformed the pathomics-only model (model 3: AUC = 0.863), whereas in ileocolonic CD, the pathomics-only model (model C: AUC = 0.866) outperformed model E. Pathomics-only models (model 3 and model C) outperformed radiomics-only models (model 1: AUC = 0.799; model A: AUC = 0.827), which in turn outperformed the radiomics + clinical models (model 4: AUC = 0.731; model D: AUC = 0.687). The radiomics + clinical models (model 4 and model D) also outperformed the clinical-only models (model 2: AUC = 0.669; model B: AUC = 0.636).

As shown in Table 5, DeLong tests confirmed significant differences in AUCs. In ileal CD, model 5 and model 7 outperformed model 2 (P = 0.029 and P = 0.001, respectively), and model 6 outperformed both model 2 and model 4 (P = 0.003 and P = 0.044, respectively). In ileocolonic CD, multiple models (model A, model C, model E, model F, and model G) outperformed model B (all P < 0.05). Model F and model G also outperformed model D (P = 0.007 and P = 0.017, respectively).

Table 5 Delong test comparison between models of area under the curves in the testing set.

Models		P value
		Clinical	Radiomics	Pathomics	Radiomics and clinical	Pathomics and clinical	Radiomics and pathomics	Combined
Clinical	Ileal		0.259	0.121	0.623	0.029	0.003	0.001
	Ileocolonic		0.004	0.006	0.635	0.012	< 0.001	< 0.001
Radiomics	Ileal	0.259		0.496	0.495	0.227	0.094	0.114
	Ileocolonic	0.004		0.626	0.087	0.905	0.209	0.297
Pathomics	Ileal	0.121	0.496		0.245	0.890	0.503	0.203
	Ileocolonic	0.006	0.626		0.112	0.218	0.446	0.464
Radiomics and clinical	Ileal	0.623	0.495	0.245		0.122	0.044	0.119
	Ileocolonic	0.635	0.087	0.112		0.234	0.007	0.017
Pathomics and clinical	Ileal	0.029	0.227	0.890	0.122		0.321	1.000
	Ileocolonic	0.012	0.905	0.218	0.234		0.142	0.107
Radiomics and pathomics	Ileal	0.003	0.094	0.503	0.044	0.321		1.000
	Ileocolonic	< 0.001	0.209	0.446	0.007	0.142		0.629
Combined	Ileal	0.001	0.114	0.203	0.119	1.000	1.000
	Ileocolonic	< 0.001	0.297	0.464	0.017	0.107	0.629

DISCUSSION

This study evaluated seven prediction models for MH in ileal and ileocolonic CD. The integrated radiomics + pathomics models (model 6 for ileal CD; model F for ileocolonic CD) demonstrated the best performance in both the training and testing sets.

For ileal CD, the performance ranking was as follows: The pathomics + clinical model (model 5) marginally outperformed the pathomics-only model (model 3), model 3 outperformed the combined model (model 7), model 7 slightly outperformed the radiomics-only model (model 1), model 1 outperformed the radiomics + clinical model (model 4), and model 4 outperformed the clinical-only model (model 2).

For ileocolonic CD, the combined model (model G) outperformed the pathomics-only model (model C), model C outperformed the radiomics-only model (model A), model A marginally outperformed the pathomics + clinical model (model E), model E outperformed the radiomics + clinical model (model D), and model D outperformed the clinical-only model (model B). Overall, these findings indicate that integrated radiomics-pathomics models provide effective and reliable prediction of MH in CD.

This study identified baseline Alb as an independent clinical predictor of MH in ileal CD and FC in ileocolonic CD. Alb, synthesized in the liver, reflects both intestinal inflammatory burden and nutritional status and serves as a biomarker of CD severity[28,29]. FC a calcium/and zinc-binding protein released by neutrophils during intestinal inflammation provides a noninvasive measure of mucosal inflammation[30,31]. However, both biomarkers demonstrated limited predictive performance in the clinical models (Alb: AUC = 0.669; FC: AUC = 0.636), suggesting that reliance on a single marker for MH prediction may lead to misclassification due to limited specificity. In contrast, radiomics captures subtle pathophysiological changes by noninvasively extracting high-dimensional imaging features, enabling more robust prediction.

Radiomics represents a promising advance in CD research and has demonstrated clinical utility in recent years. Chen et al[32] developed a CTE-based radiomics model to predict IFX loss-of-response in 186 biologic-naive patients, with strong performance (AUC = 0.880). Similarly, a study of 106 treatment-naive patients established a CTE-derived nomogram that combined radiomic and clinical features to predict MH (AUC = 0.877)[16]. Building on this work, we developed radiomics models to predict MH by selecting representative ileal and ileocolonic segments on CTE. Our models achieved AUCs of 0.799 (ileal) and 0.827 (ileocolonic).

Although radiomics captures macroscopic imaging patterns, it cannot characterize molecular-level pathophysiology. Pathomics helps address this limitation by quantifying histologic features that may reflect underlying genetic and molecular mechanisms. Accordingly, we incorporated pathomic feature extraction from biopsy specimens to provide complementary biological insights.

Pathomics has emerged as a promising field and has been applied primarily in oncology for prognostic prediction (renal cell carcinoma), diagnostic classification (bladder cancer), survival analysis (bladder cancer), and assessment of response to neoadjuvant therapy (rectal cancer)[17,18,33]. In this study, pathomics models developed from histopathologic features of ileal and ileocolonic lesions achieved AUCs of 0.863 and 0.866, respectively, outperforming the corresponding radiomics models. This finding is consistent with that of Wang et al[34], who reported that integrated pathomics-radiomics modeling improved accuracy for predicting lung metastasis from colorectal cancer compared with single-modality approaches. Our integrated radiomics + pathomics models demonstrated superior performance in both ileal (AUC = 0.915) and ileocolonic (AUC = 0.922) CD, outperforming all single-modality and other combined models. This synergy likely reflects complementary information, with CTE-derived structural heterogeneity combined with histopathologic cellular heterogeneity providing more comprehensive biomarkers for MH prediction.

Notably, incorporating clinical factors reduced performance (ileal combined AUC = 0.812; ileocolonic combined AUC = 0.908). Although the ileocolonic combined model achieved performance comparable to the radiomics + pathomics model, the ileal combined model showed high sensitivity (0.944) and specificity (0.769) despite a lower AUC.

This study has several limitations. First, the single-center retrospective design and relatively small sample size, particularly after stratification into ileal and ileocolonic subgroups, may introduce selection bias and limit the generalizability of the findings. Although the sample size met the basic requirements for model development, a larger cohort would further improve the stability and generalizability of complex models using high-dimensional features. To facilitate clinical translation, future validation studies should adopt a multicenter prospective design with a larger sample size and include patients with isolated colonic involvement to enable a more comprehensive evaluation. Second, pathomics feature extraction was limited to representative lesion sites rather than all affected intestinal segments. The corresponding CTE-based radiomics analysis was similarly restricted. We plan to expand histopathologic sampling to include all diseased segments to enable a more comprehensive analysis. Third, semi-automatic segmentation of intestinal lesions on CTE may introduce variability in radiomics feature extraction. Future work will focus on developing a fully automated segmentation algorithm to capture pathology across the entire intestinal wall.

CONCLUSION

The baseline radiomics-pathomics integrated models demonstrated robust predictive accuracy for MH at 6-9 months after IFX therapy in patients with ileal and ileocolonic CD.

ACKNOWLEDGEMENTS

The authors would like to thank research centers (the First Affiliated Hospital of Anhui Medical University and Anhui Medical University) for their contributions.