Two parallel systems constructed from complementary optical coherence tomography and histopathology for the assessment of murine intestinal inflammation: a dual-modality assessment

Abstract

Background

Quantitative assessment of intestinal inflammation among patients with inflammatory bowel disease (IBD) remains a clinical challenge. Conventional histopathology lacks robust quantification, and a real-time, noninvasive alternative is needed. This study aimed to establish a quantitative histopathological system for IBD and to develop a corresponding optical coherence tomography (OCT) system for real-time detection.

Methods

Murine models for both ulcerative colitis (UC) and Crohn disease (CD) were developed, and corresponding intestinal histopathological samples were obtained. Artificial intelligence-based image segmentation enabled systematic analysis of pathological images, including two-dimensional planar structures, three-dimensional volumetric morphology, texture features, and grayscale histograms, through which a comprehensive histopathological evaluation system for precise quantitative assessment of intestinal inflammation severity was established. Based on this framework, an OCT system was developed for intestinal wall structural analysis, which demonstrated high concordance with histopathological benchmarks while enabling real-time, noninvasive quantitative detection.

Results

Both systems effectively differentiated UC, CD, and normal tissue. Planar histopathology revealed significant differences in these models in comparison to controls. The median mucosal fold angle was markedly elevated in UC [median 63.854°, interquartile range (IQR), 37.110–91.151°] and CD (49.925°, IQR, 34.140–62.550°) mice as compared controls (7.635°, IQR, 5.975–10.873°) (both P values <0.001). Submucosal area and cell count were higher in both disease groups, while the gland count was lower (all P values <0.05). Mucosal area ratio decreased from 73.608% in controls to 66.703% in UC mice (P=0.020) and to 59.182% in CD mice (P<0.001), whereas the submucosa ratio increased from 8.127% to 10.812% (P=0.031) and 16.076% (P<0.001), respectively. Three-dimensional reconstruction confirmed increased intestinal wall volumes in both disease groups (all P values <0.05). OCT texture analysis showed decreased energy and increased entropy in the UC and CD groups as compared to the control group (P<0.05), indicating disrupted microstructure. Gray-scale analysis revealed a shift toward lower values in diseased tissues (P<0.001). Pathological texture analysis demonstrated significantly elevated contrast in CD mice (median 153.425) as compared with UC (62.160) and control (70.495) mice (P<0.001), suggesting that CD and UC each have distinct textural signatures.

Conclusions

This work establishes an integrated framework: a quantitative histopathological system providing novel diagnostic indices, and a complementary OCT system enabling real-time assessment. It offers a proof of concept for multimodal quantitative biomarkers in IBD, an objective tool for preclinical research, and a foundation for future OCT endoscopic applications in clinical evaluation.

Introduction

Histopathological assessment remains the gold standard for diagnosing and monitoring disease activity among patients with inflammatory bowel disease (IBD) (1). However, conventional pathology is inherently limited by its reliance on subjective interpretation, substantial interobserver variability, and the lack of accurate and reproducible quantitative metrics (2). In addition, the histological manifestations of IBD, including the degree and spatial distribution of inflammatory cell infiltration and glandular architectural distortion, are often diffuse and heterogeneous (3), posing significant challenges to objective and standardized evaluation.

Recent advances in digital pathology and artificial intelligence (AI) have facilitated the automated, high-throughput, and quantitative assessment of intestinal inflammation (4). To further overcome the subjectivity of conventional assessment, label-free quantitative imaging techniques have emerged as promising alternatives. Notably, digital holographic microscopy (DHM) and quantitative phase imaging (QPI) have convincingly demonstrated their value in the objective quantification of intestinal inflammation in both experimental colitis and UC patients (5,6). However, these techniques primarily acquire information in the horizontal plane of the scanned tissue, which may not fully capture the architectural alterations related to depth. Furthermore, such surface-based imaging and histopathology are invasive and constrained by focal tissue sampling, which may not accurately represent the overall disease state (7). To address these limitations, we conducted a study using optical coherence tomography (OCT), a high-resolution imaging modality often described as “optical biopsy” due to its ability to visualize tissue microstructure in real time and approach the level of detail observed in conventional histology (8). Critically, unlike DHM/QPI, two-dimensional OCT (2D-OCT) primarily captures longitudinal (cross-sectional) information, and three-dimensional OCT (3D-OCT) further reconstructs comprehensive 3D volumetric information, offering unique insights into the tissue’s depth and structural integrity.

The aims of this study were to integrate histopathological and OCT data to identify novel multidimensional quantitative biomarkers and to develop and validate a comprehensive AI-driven framework for their evaluation. The overarching objective was to establish a methodological blueprint for precise IBD stratification, treatment monitoring, and drug development, thereby facilitating future diagnostic advances in OCT optical endoscopy (9,10). We present this article in accordance with the ARRIVE reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0232/rc).

Methods

Animal models and tissue preparation

Induction of murine colitis models

Specific pathogen-free female C57BL/6 mice (6–8 weeks old) were obtained from Wu’s Animal Center (Fuzhou, China) and housed at Fujian Medical University under controlled conditions (24 ℃, a 12-hour light-dark cycle, and free access to food and water) (11). After 1 week of acclimatization, 103 mice were randomly allocated into three experimental groups. The randomization was performed with a computer-generated random number sequence. Group allocation was based solely on the random sequence, and investigators were blinded to the allocation during the modeling period. The control group (n=18) received a standard chow diet without additional treatment. The dextran sulfate sodium (DSS)-induced ulcerative colitis (UC) group (n=47) was administered 2.5% DSS in drinking water for 7 days, which was followed by a 14-day recovery period for three consecutive cycles to induce chronic UC (12,13). The 2,4,6 trinitrobenzene sulfonic acid (TNBS)-induced Crohn disease (CD) group (n=38) received weekly intrarectal administration of 5% TNBS in 75% ethanol at a dose of 0.04 mL per 10 g body weight for 9 consecutive weeks, with the first induction being preceded by a 24-hour fasting period (14).

At the end of the 9-week modeling period, mice were euthanized by cervical dislocation, and distal colon tissues were harvested for subsequent imaging and histological analyses. All animal procedures were approved by the Institutional Animal Care and Use Committee of Fujian Medical University (approval No. IACUCFJMU2024-Y-2494) and were conducted in accordance with institutional guidelines for the care and use of laboratory animals.

Tissue collection and histological processing

Following tissue harvest, all specimens were obtained from the distal colon at a standardized site: a 0.5-cm segment located 2 cm proximal to the anal margin. These segments were cut open longitudinally, rinsed with saline, and flattened on glass slides after being dried on filter paper to ensure a consistent orientation for both OCT imaging and subsequent histology. Colon samples from the control (n=18), DSS-induced UC (n=47), and TNBS-induced CD (n=38) groups underwent OCT imaging prior to paraffin embedding (Figure 1A,1B). Of these, a subset of 21 specimens (control: n=8; UC: n=6; CD: n=7) was subsequently used for 3D pathological reconstruction, and the OCT volumetric data acquired from these matched specimens were used for corresponding analyses. Consecutive tissue sections were then prepared via interval slicing, with every other section collected and stained with hematoxylin and eosin (HE) for planar analysis, 3D reconstruction, texture feature extraction, and gray-scale analysis. This process yielded 50 HE-stained sections per specimen, covering a total length of 3000 µm, which preserved tissue continuity for 3D reconstruction while optimizing resource utilization. This approach balanced spatial resolution with experimental requirements, with an emphasis on overall morphological assessment rather than fine structural details (15).

Figure 1.

Schematic diagram of the 2D- and 3D-OCT images and histopathological validation. (A) 2D-OCT image. (B) 3D-OCT image. (C) UC model induced by DSS: (a) representative HE and Masson’s trichrome staining of colonic tissues from the water and DSS groups, (b) quantification of collagen fiber content, (c) Sirius Red staining observed under bright field and polarized microscopy shown alongside HE/Masson’s trichrome staining, and (d) quantification of the polarization-positive area. (D) CD model induced by TNBS: (a) representative HE and Masson’s trichrome staining of colonic tissues from the water and TNBS groups, (b) quantification of collagen fiber content, (c) Sirius Red staining with HE/Masson’s trichrome, and (d) quantification of the polarization-positive area. *, P<0.05; **, P<0.001. All histopathological staining (H&E, Masson’s trichrome, and Sirius Red) was imaged at 200× magnification. Scale bar =50 μm. 2D, two-dimensional; 3D, three-dimensional; CD, Crohn disease; DSS, dextran sulfate sodium; HE, hematoxylin and eosin; OCT, optical coherence tomography; TNBS, 2,4,6-trinitrobenzenesulfonic acid; UC, ulcerative colitis.

The construction of the animal disease models was validated through histological assessment of murine colonic tissue samples via HE staining, Sirius Red staining, and Masson’s trichrome staining (Figure 1C,1D) (16).

Multimodal image acquisition

Digital scanning of pathology slides

The HE-stained slides prepared in the above-described procedure were digitized to generate high-resolution pathological images for subsequent computational analysis (17). During image analysis, different magnification scales were applied according to the specific feature extraction tasks. Planar geometric vector data were acquired at 40× magnification, whereas cell counting and gland quantification were performed at 200× magnification.

OCT imaging system and protocol

A custom-built frequency-domain OCT instrument with a central wavelength of 1,000 nm, an axial resolution of 4 µm, and a scanning frequency of 16 kHz was employed in this study (18). In this system, the light beam was evenly divided into the sample arm and the reference arm of a fiber-based Michelson interferometer. The sample beam was guided by galvanometer-actuated mirrors to enable bidirectional scanning. After being focused on the sample, the light reflected from the tissue interfered with the reference beam within the interferometer. This interference produced a frequency clock signal with evenly spaced optical frequencies, which was used for OCT signal sampling (19).

With this system, 2D-OCT images were acquired from three experimental groups: the control group (n=18), the DSS-induced UC group (n=47), and the TNBS-induced CD group (n=38). Subsequently, 3D images were obtained via volumetric scanning. After standardized quality control procedures were completed, the final 3D-OCT analysis included 45 qualified datasets, comprising 12 normal controls, 15 UC samples, and 18 CD samples. Volumetric scanning was performed with a 5-µm step size along the y-axis over a total scanning range of 3,000 µm, generating 600 x-y plane and 500 x-z plane cross-sectional images per specimen (20).

To ensure anatomical correspondence between OCT volumes and HE sections, the OCT scanning region was precisely matched to the same 0.5-cm distal colonic segment that was subsequently processed for histology. The OCT en face imaging plane was aligned with the longitudinal axis of the flattened tissue, and the first HE section was prepared from the OCT-imaged surface to enable direct cross-modal comparison. For OCT-to-histology registration, OCT en face images and HE planar images were manually co-registered via intrinsic anatomical landmarks—including but not limited to crypt openings and vascular patterns—as fiducial markers.

AI-driven image segmentation

Annotation of pathological images and dataset preparation

Pathological images were annotated with LabelMe software to generate training datasets (Figure 2A). Image preprocessing included normalization of pixel values to the range of 0 to 1, resizing, and data augmentation to reduce the risk of overfitting. All annotations were independently verified by two senior physicians and one pathologist, and strong interrater agreement was achieved, with a κ value of 0.92 and a 95% confidence interval of 0.90 to 0.94 (21).

Figure 2.

Workflow of the artificial intelligence-powered segmentation and morphological analysis. (A) Representative pathological images with segmentation masks generated by the U-Net model. Quantitative metrics derived from segmentation including the mucosal area, gland area, gland count, and gland area ratio (%). (B) Demonstration of cell counting in pathological images, with C1 indicating the mucosal cell count (11,491.57/mm²) and C2 indicating the submucosal cell count (1,178.48/mm²). (C) Workflow for pathological image segmentation via the U-Net model for tissue layers: (a) mucosal layer, (b) submucosal layer, and (c) muscular layer. (D) Workflow for OCT image segmentation via the U-Net model for tissue layers: (a) mucosa, (b) submucosa, and (c) muscularis. (E) Illustration of 2D planar parameters derived from pathology images, including representative measurements for layer thickness, area, and ratios. (F) Representative OCT images showing segmentation of tissue layers and quantitative measurements of layer thickness, such as mucosal fold thickness and submucosal thickness. Representative thickness measurements are indicated. All histopathological stains (H&E) were imaged at 40× magnification; scale bar =200 μm. 2D, two-dimensional; OCT, optical coherence tomography.

U-Net architecture and training for gland segmentation

A U-Net architecture was implemented for the segmentation of glandular ducts (Figure 2A). To prevent data leakage, the dataset was divided at the specimen (mouse) level rather than at the image level, ensuring that all sections from the same specimen were assigned exclusively to either the training set or the validation set. Specifically, 90% of the specimens were used for training and the remaining 10% for validation. The model was trained for 200 epochs with a batch size of 8 via the Adam optimizer. The dataset was divided into a training set comprising 90% of the images and a validation set comprising the remaining 10%. The model was trained for 200 epochs with a batch size of 8 via the Adam optimizer. The initial learning rate was set to 1×10⁻⁴ and was reduced by 50% every 100 epochs. The model was trained with a weighted combination of cross-entropy loss (L) (Eq. [1]) and Dice coefficient (Eq. [2]) as the objective function, defined as follows: L=0.5·L_CE + 0.5·L_Dice, where L_CE denotes the cross-entropy loss and L_Dice denotes the Dice loss.To mitigate overfitting, data augmentation, dropout, and early stopping were applied, with training terminated when validation performance plateaued (22). Optimization was performed via the Adam optimizer with a dynamic learning rate adjustment strategy (Eq. [3]). As shown in Figure S1, the training and validation loss curves exhibited stable convergence with no significant overfitting, and the Dice coefficients steadily increased and plateaued, confirming the effectiveness of the chosen loss weighting and regularization strategies.

$${\mathrm{Loss}}_{ce}=-\frac{1}{N}\sum _{i=1}^N[y_i\cdot \log ({\hat{y}}_i)+(1-y_i)\cdot \log (1-{\hat{y}}_i)]$$ [1]

Eq. [1] defines the cross-entropy loss function, which measures the difference between model predictions and true labels.

$$\mathrm{Dice}=\frac{2|A\cap B|}{|A\cup B|}$$ [2]

Eq. [2] defines the Dice coefficient for measuring segmentation overlap with the ground truth.

$${\eta }_{t+1}=\frac{{\eta }_t}{1+\lambda t}$$ [3]

Eq. [3] formulates the dynamic learning rate adjustment for stable parameter optimization. λ is the decay factor (hyperparameter) that determines the magnitude of the drop in the learning rate. A larger λ accelerates decay. t represents the current training iteration (epoch/step). As training progresses, t monotonically increases → λ*t grows steadily.

Because of the limited number of specimens, an independent held-out test set was not established. Therefore, final model performance was evaluated on the validation set, with a Dice coefficient of 0.751 being achieved (Figure S1). The resulting binary segmentation masks were subsequently used for automated nuclear counting to enable quantitative cellular analysis (23) (Figure 2B).

All models were developed with Python version 3.8 (Python Software Foundation, Wilmington, DE, USA) with PyTorch version 1.3.1 and torchvision version 0.4.2. Numerical computations and image processing were performed via numpy version 1.17.4 and the Python version of opencv 4.1.2.30, respectively (24).

U-Net architecture and training for tissue layer segmentation

A U-Net-based segmentation pipeline was developed for the accurate delineation of mucosal, submucosal, and muscularis layers in histopathological images (Figure 2C). The same method was applied to segment these layers in the corresponding OCT images (Figure 2D). The network architecture was identical to that used for glandular duct segmentation. Annotations for tissue layer segmentation were independently verified, yielding substantial interrater agreement with a κ value of 0.707 and a 95% confidence interval of 0.663 to 0.750 (25). Following segmentation, the output masks were used to calculate 22 geometric parameters, including layer-specific thicknesses and areas for the mucosa, submucosa, and muscularis, as well as structural ratios and angular measurements of mucosal folds (Figure 2E). To enable direct cross-modal comparison, the same analytical pipeline was applied to the corresponding OCT images (Figure 2F). This standardized approach allowed for the comprehensive morphometric characterization of the inflammatory burden across both imaging modalities.

Extraction of quantitative multidimensional features

Planar morphometric analysis

Following AI-driven segmentation of tissue layers in both pathological and OCT images, algorithms were developed for contour detection, area calculation, and angular measurement of mucosal folds. This allowed for the computation of 22 geometric parameters, including layer-specific thicknesses (e.g., mucosal fold thickness and planar submucosal thickness), area measurements (e.g., submucosal area and muscularis area), and structural ratios (e.g., mucosal area ratio) (26).

Three-dimensional volumetric reconstruction and analysis

A novel bucket model was developed for 3D reconstruction through the conceptualization of volumetric data as stacked 2D layers. This unified framework sequentially analyzed pathological sections and OCT layers and integrated them into complete 3D structures with preserved spatial relationships for stereoscopic analysis (Figure 3) (27). For histopathological analysis, a standardized reconstruction protocol was applied that integrated skip slice histology with a section thickness of 5 µm and an interslice interval of 5 µm (Figure 3A), together with OCT imaging at a thickness of 5 µm (Figure 3B). Nuclear counts were calculated from doubled 2D counts (Eq. [4]), while glandular ducts were quantified with Open Source Computer Vision (OpenCV)-based diameter-adjusted algorithms (Eq. [5]). Specifically, the total glandular duct count was estimated as follows: total gland count = Σ(10/D_i), where D_i denotes the diameter of the i-th glandular duct (in µm). This inverse scaling requires the assumption that the number of ducts is proportional to 1/D_i, with 10 (µm) serving as the reference normalization factor corresponding to the standard duct diameter, consistent with stereological correction principles for tubular structures (28,29). Volumetric data were generated through thickness-scaled layer-wise area summation, with a slice thickness of 10 µm for pathology and 5 µm for OCT. This was followed by unit conversion (Eqs. [6,7]), which ensured consistency in cross-modal quantification (27). Specifically, the volume of the pathological target region was calculated as follows: volume = (Σ A_i)×10/(10×9) (mm³), where A_i denotes the total area of the target region in the i-th slice (in mm²). This layer-wise summation approach follows the classical Cavalieri principle for volume estimation, a well-established stereological technique extensively validated in design-based stereology (28,29).

Figure 3.

Three-dimensional reconstruction and quantitative volumetric modeling. (A) Three-dimensional parameters derived from pathological images, including mucosal cell count (106/mm³), submucosal cell count (106/mm³), gland count (106/mm³), gland volume percentage, mucosal volume (106/mm³), submucosal volume (106/mm³), muscular volume (106/mm³), intestinal wall volume (106/mm³), mucosal volume ratio, submucosal volume ratio, and muscular volume ratio. (B) Three-dimensional parameters derived from OCT imaging, including mucosal volume (106/mm³), submucosal volume (106/mm³), muscular volume (106/mm³), intestinal wall volume (106/mm³), mucosal volume ratio, submucosal volume ratio, and muscular volume ratio. All histopathological stains (H&E) were imaged at 200× magnification; scale bar =50 μm. OCT, optical coherence tomography.

$$\mathrm{TotalCellCount}=2\times \sum _{i=1}^n{C}_i$$ [4]

This equation was used to calculate the total cell count in each pathological tissue layer, where Cᵢ represents the cell count in the i-th image.

$$\mathrm{TotalGlandCount}=\sum _{i=1}^n\frac{10}{D_i}$$ [5]

This equation estimates the total number of glandular ducts based on duct diameter, where Dᵢ denotes the diameter of the i-th glandular duct.

$$V=\left(\sum _{i=0}^n{A}_i\right)\ast {10}^{-8}({\mathrm{mm}}^3)$$ [6]

In this equation, V is the total volume of the pathological target region, and Aᵢ denotes the total area of the target region in the i-th slice.

$$\mathrm{OCTVol}=5\times \left(\sum _{i=1}^N{A}_i\right)\times {10}^{-9}(\mathrm{m}{\mathrm{m}}^3)$$ [7]

In this equation, OCTVol is the volume of the target region derived from OCT images, and Aᵢ denotes the area of the target region in the i-th OCT image.

Image texture analysis with gray-level co-occurrence matrix

OCT images were directly analyzed for texture parameters because they are inherently gray scale (Figure 4A,4B). The gray-scale mean was calculated by dividing the sum of gray-scale values by the total number of pixels. Gray level co-occurrence matrix-derived texture parameters, including energy, entropy, contrast, correlation, and homogeneity, were quantified for the global OCT structure (Figure 4B,4C). Calculations were performed at two pixel distances (d=[1, 2]) and four directions (θ=0º, θ=45º, θ=90º, and θ=135º) to comprehensively characterize textural information.

Figure 4.

Image texture and gray-scale feature analysis. (A) OCT image and pathological image. (B) GLCM: (a) energy equation, (b) entropy equation, (c) contrast equation, (d) correlation equation, and (e) homogeneity equation. (C) OCT image texture parameters and pathological image texture parameters. (D) Representative OCT 2D image, identical to that shown in Figure 1A, used for 2D visualization and gray-scale histogram analysis. (E) Conversion of pathological images to gray scale via the luminance method, followed by gray-scale histogram generation. All histopathological stains (H&E) were imaged at 40× magnification; scale bar =200 μm. 2D, two-dimensional; GLCM, gray-level co-occurrence matrix; OCT, optical coherence tomography.

For pathological images, color images were first converted to gray scale via the luminance method (Figure 4A). Subsequently, the gray-level co-occurrence matrix method was applied for texture feature extraction (Figure 4B,4C) (30).

Gray-scale histogram analysis

OCT gray-scale images with low intralayer contrast were analyzed according to gray-scale histogram distributions. The 8-bit grayscale range from 0 to 255 was divided into five equal intervals (0–49, 50–99, 100–149, 150–199, and 200–255) via fixed thresholds. Pixel distributions within each interval were quantified with the OpenCV cv2.calcHist function to compare inflammatory tissues with normal tissues (Figure 4D). The same analytical approach was applied to murine intestinal pathology images, allowing for the evaluation of the gray-scale distribution across full-thickness intestinal wall layers within the predefined intervals (Figure 4E) (31).

Statistical analyses and evaluation of diagnostic performance

Statistical analyses were conducted with SPSS version 27.0 (IBM Corp., Armonk, NY, USA), and graphical visualizations were generated via GraphPad Prism (Dotmatics, Boston, MA, USA). For continuous variables, normality was initially assessed. Normally distributed data are expressed as the mean ± standard deviation and were compared with the independent samples t-test, whereas nonnormally distributed data are expressed as the median and interquartile range (IQR) and were compared with the Mann-Whitney U test. Categorical variables were analyzed via the chi-squared test, and correlations were assessed through Spearman rank correlation analysis. A two-sided P value of less than 0.05 was considered statistically significant.

Interrater agreement between the automated segmentation results and ground truth annotations was quantitatively evaluated via the Cohen kappa coefficient (32).

Receiver operating characteristic (ROC) curves were used to assess the diagnostic performance of individual parameters and the integrated model, with an area under the curve (AUC) >0.60 considered indicative of discriminative ability (95% confidence interval: 0.52–1.00) (33). For construction of the integrated multimodal model, features derived from planar morphometric analysis, texture analysis, and gray-scale histogram analysis were used as inputs for a logistic regression classifier. Although 3D volumetric parameters demonstrated promising discriminative ability when evaluated individually, they were excluded from the primary model to mitigate overfitting risks given the limited sample size. To prevent information leakage and ensure unbiased performance estimation, all preprocessing steps—including normalization and feature selection—and model training were performed strictly within each iteration of the leave-one-out cross-validation (LOOCV) framework. Specifically, for each LOOCV iteration, normalization parameters (mean and standard deviation) were estimated from the training set and applied to the corresponding single validation sample. Feature selection was conducted exclusively within the training set, and the selected feature subset was then applied to the validation sample. The final model was evaluated through rigorous LOOCV across the entire dataset of 103 qualified samples, comprising 18 controls, 47 UC cases, and 38 CD cases (34).

Results

Establishment and pathological validation of the IBD models

To confirm the successful establishment of the IBD models, comprehensive histological evaluation of murine colonic tissues was performed. Through comparison with the control group, it was found that both the DSS-induced and TNBS-induced models successfully recapitulated the characteristic pathological features of UC and CD, respectively. These features included pronounced inflammatory cell infiltration and tissue remodeling (Figure 1C,1D).

The DSS-induced UC model exhibited histopathological alterations predominantly localized to the mucosa and submucosa (Figure 1C, panel a). Hallmark features included extensive inflammatory cell infiltration, disruption of crypt architecture, depletion of goblet cells, and mucosal ulceration. Masson’s trichrome and Sirius Red staining further demonstrated a marked increase in collagen deposition within the lamina propria and submucosa (Figure 1C, panels a and c). In the comparison with the control group, quantitative analysis confirmed significant increases in collagen fiber content (P<0.001) and polarization-positive area (P<0.05) (Figure 1C, panels b and d).

In contrast, the TNBS-induced CD model exhibited transmural inflammation involving all layers of the intestinal wall (Figure 1D, panel a). This was characterized by extensive inflammatory infiltrates extending through the submucosa into the muscularis layer. Similar to the UC model, the TNBS-induced CD model exhibited a substantial increase in fibrosis, as evidenced by Masson’s trichrome and Sirius Red staining (Figure 1D, panels a and c). Quantitative analysis further confirmed significant elevations in collagen fiber content (P<0.001) and polarization-positive area (P<0.05) in comparison with controls (Figure 1D, panels b and d).

Comparative planar 2D morphometric analysis of pathological and OCT images

Following AI-driven segmentation, quantitative analysis of 2D planar images revealed significant and consistent morphological alterations in both the UC and CD groups as compared with the control group.

Pathological planar analysis (Table 1) showed that, compared with the control group (n=18), both the UC (n=47) and CD (n=38) groups showed increases in the median mucosal fold angle (control: 7.635°, IQR, 5.975–10.873; UC: 63.854°, IQR, 37.110–91.151°, P<0.001; CD: 49.925°, IQR, 34.140–62.550°, P<0.001), mucosal fold thickness (control: 219.738 μm, IQR, 197.146–244.638 μm; UC: 250.954 μm, IQR, 215.512–305.268 μm, P=0.021; CD: 260.470 μm, IQR, 222.290–319.910 μm, P=0.001), submucosal area (control: 141,725.720 μm², IQR, 111,924.014–220,571.660 μm²; UC: 203,093.919 μm², IQR, 136,683.270–299,696.373 μm², P=0.034; CD: 368,237.832 μm², IQR, 272,441.008–504,929.621 μm², P<0.001), and submucosal area ratio (control: 8.127%, IQR, 6.829–10.505%; UC: 10.812%, IQR, 8.190–13.935%, P=0.031; CD: 16.076%, IQR, 12.633–19.456%, P<0.001). Meanwhile, relative to the control group, the UC and CD groups had a lower median mucosal area ratio (control: 73.608%, IQR, 63.836–75.771%; UC: 66.703%, IQR, 58.513–73.382%, P=0.020; CD: 59.182%, IQR, 52.992–65.196%, P<0.001), gland count (control: 107.465×10⁶ μm², IQR, 78.457–204.287 ×10⁶ μm²; UC: 50.952×10⁶ μm², IQR, 38.454–80.125 ×10⁶ μm², P<0.001; CD: 87.049×10⁶ μm², IQR, 69.894–103.841 ×10⁶ μm², P=0.042), and gland area ratio (control: 58.495%, IQR, 44.198–90.803%; UC: 41.916%, IQR, 29.985–56.910%, P=0.012; CD: 41.413%, IQR, 33.784–52.559%, P=0.011). Submucosal cell count was elevated in both disease groups (control: 1,568.845×10⁶ μm², IQR, 1,038.539–2,434.193 ×10⁶ μm²; UC: 3,316.827×10⁶ μm², IQR, 2,288.058–5,804.423 ×10⁶ μm², P<0.001; CD: 3,757.601×10⁶ μm², IQR, 2,339.072–5,742.869 ×10⁶ μm², P<0.001).

Table 1. Comparison of pathological plane parameters between the UC, CD, and control groups.

Category	Parameter	Control group (n=18)	UC group (n=47)	CD group (n=38)	Control vs. UC		Control vs. CD		UC vs. CD
					Z/T	P	Z/T	P	Z/T	P
Pathological plane geometry parameter	Angle of mucosal fold (°)	7.635 (5.975, 10.873)	63.854 (37.110, 91.151)	49.925 (34.140, 62.550)	−6.040	<0.001	−5.509	<0.001	−2.360	0.018
	Thickness of mucosal fold (um)	219.738 (197.146, 244.638)	250.954 (215.512, 305.268)	260.470 (222.290, 319.910)	−2.302	0.021	−3.246	0.001	−0.831	0.406
	Submucosal area (um2)	141,725.720 (111,924.014, 220,571.66)	203,093.919 (136,683.270, 299,696.373)	368,237.832 (272,441.008, 504,929.621)	−2.126	0.034	−5.070	<0.001	−4.579	<0.001
	Mucosal area ratio (%)	73.608 (63.836, 75.771)	66.703 (58.513, 73.382)	59.182 (52.992, 65.196)	−2.331	0.020	−4.404	<0.001	−3.297	<0.001
	Submucosa area ratio (%)	8.127 (6.829, 10.505)	10.812 (8.190, 13.935)	16.076 (12.633, 19.456)	−2.155	0.031	−4.509	<0.001	−3.597	<0.001
Cell and glandular tube count in pathological planes	Submucosal cell count/106 (μm2)	1,568.845 (1038.539, 2,434.193)	3,316.827 (2,288.058, 5,804.423)	3,757.601 (2,339.072, 5,742.869)	−3.401	<0.001	−3.474	<0.001	−0.247	0.805
	Gland count/106 (μm2)	107.465 (78.457, 204.287)	50.952 (38.454, 80.125)	87.049 (69.894, 103.841)	−4.325	<0.001	−2.035	0.042	−3.924	<0.001
	Gland area ratio (%)	58.495 (44.198, 90.803)	41.916 (29.985,56.910)	41.413 (33.784, 52.559)	−2.507	0.012	−2.544	0.011	−0.199	0.842
OCT plane geometry parameter	Angle of mucosal fold (°)	22.910 (18.570, 28.930)	68.230 (45.460, 85.335)	58.150 (42.670, 94.160)	−5.571	<0.001	−5.562	<0.001	−0.345	0.730
	Thickness of mucosal fold (μm)	166.639±46.638	210.3519±64.550	203.431±64.311	−2.618	0.011	−2.168	0.035	0.492	0.624
	Submucosal area (μm2)	343,616.678 (26,360.722, 450,214.902)	499,315.101 (403,886.629, 604,158.708)	582,836.576 (491,965.488, 712,909.746)	−3.533	<0.001	−4.982	<0.001	−2.440	0.015

Data following a normal distribution are presented as the mean ± SD (analyzed via the t test); nonnormally distributed data (unmarked) are presented as the median (25th, 75th percentiles) (analyzed via the Mann-Whitney U test). P<0.05 indicates a statistically significant difference. CD, Crohn disease; OCT, optical coherence tomography; SD, standard deviation; UC, ulcerative colitis.

OCT planar analysis (Table 1) yielded consistent findings, with both the disease groups demonstrating increases in median mucosal fold angle (control: 22.910°, IQR, 18.570–28.930°; UC: 68.230°, IQR, 45.460–85.335°, P<0.001; CD: 58.150°, IQR, 42.670–94.160°, P<0.001), mucosal fold thickness (control: 166.639±46.638 μm; UC: 210.352±64.550 μm, P=0.011; CD: 203.431±64.311 μm, P=0.035), and submucosal area (control: 343,616.678 μm², IQR, 26,360.722–450,214.902 μm²; UC: 499,315.101 μm², IQR, 403,886.629–604,158.708 μm², P<0.001; CD: 582,836.576 μm², IQR, 491,965.488–712,909.746 μm², P<0.001). A complete comparison of all parameters is provided in Table S1.

Correlation analyses (Tables S2-S5) examined the associations between these parameters. In the UC model, pathological images showed that mucosal fold angle had the strongest positive correlation with overall disease severity (r=0.755; P<0.001). Mucosal fold angle, mucosal fold thickness, submucosal area, submucosal area ratio, and submucosal cell count differed significantly between disease groups and the control group (Table 1).

Comparative 3D volumetric analysis revealed amplified tissue alterations

To address the inherent limitations of single 2D sections and to capture a more comprehensive representation of intestinal inflammation, 3D volumetric reconstruction and analysis were performed for both pathological and OCT datasets (Figure 3).

Compared with the control group, both the UC and CD groups exhibited significant volumetric alterations as quantified through 3D reconstruction. Histopathological analysis confirmed significant increases in mucosal volume, submucosal volume, and total bowel wall volume, accompanied by an increased submucosal cell count and a decreased gland count (P<0.05).

Consistent with these pathological observations, 3D-OCT analysis also demonstrated significantly increased volumes across the mucosal, submucosal, muscular, and total intestinal wall layers in both disease groups (P<0.05). The principal results of the 3D volumetric comparisons are summarized in Table 2, and the complete dataset is provided in Table S6. To further evaluate whether these group-level trends reflected genuine sample-by-sample agreement, direct Spearman rank correlation and Bland–Altman analyses were performed with matched OCT and histopathological volumetric data from the same specimens (n=21; 8 healthy control mice, 6 UC mice, and 7 CD mice) (Tables S7,S8). Given the limited sample size, the primary analyses were conducted on the pooled cohort without subgroup stratification. As shown in Figure S2, Spearman correlation revealed significant positive correlations for all four parameters: mucosal volume (ρ=0.721; P<0.001), submucosal volume (ρ=0.647; P=0.002), muscular volume (ρ=0.487; P=0.025), and intestinal wall volume (ρ=0.944; P<0.001). Bland-Altman analysis (Figure S3) confirmed clinically acceptable agreement, with ≥95% of data points falling within the 95% limits of agreement for all layers. A systematic positive bias was observed, with OCT yielding larger volumes than histopathology (mean differences: mucosal, 2.630 mm³; submucosal, 0.750 mm³; muscular, 1.490 mm³; total wall, 4.870 mm³). Significant proportional bias was noted for all parameters (all P values <0.001), indicating that discrepancies increased with tissue volume—consistent with the greater challenge of delineating boundaries in larger, inflamed specimens. Collectively, these volumetric parameters demonstrated strong potential as individual biomarkers for UC and CD, and the good-to-excellent agreement further supports OCT as a valid surrogate for histopathological volumetry.

Table 2. Three-dimensional comparison of pathological and OCT between the UC, CD, and control groups.

Category	Parameter	Control group (n=8)	UC group (n=6)	CD group (n=7)	Control vs. UC		Control vs. CD		UC vs. CD
					Z	P	Z	P	Z	P
Pathological cell and glandular duct counts	Submucosal cell count (106/mm3)	1.088 (0.925, 1.209)	6.277 (4.509, 12.221)	2.711 (2.171, 3.658)	−3.098	0.002	−3.24	0.001	−1.857	0.063
	Gland count (/mm3)	3,040 (2,093.473, 4,025.381)	308.282 (294.346, 634.479)	1278.658 (482.866, 2265.351)	−3.098	0.002	−2.315	0.021	−1.714	0.086
Three-dimensional volume of pathological geometrical parameter	Mucosal volume (mm3)	0.311 (0.237, 0.449)	0.665 (0.497, 0.679)	0.624 (0.489, 0.723)	−2.711	0.007	−2.546	0.011	−0.429	0.668
	Submucosal volume (mm3)	0.067 (0.061, 0.100)	0.123 (0.096, 0.309)	0.129 (0.111, 0.214)	−2.195	0.028	−2.43	0.015	−0.286	0.775
	Intestinal wall volume (mm3)	0.495 (0.464, 0.709)	0.897 (0.743, 1.174)	0.932 (0.807, 1.300)	−2.324	0.02	−2.43	0.015	−0.286	0.775
OCT three-dimensional parameter	Mucosal volume (mm3)	2.727 (2.363, 3.771)	4.604 (3.401, 5.441)	4.188 (3.404, 4.403)	−2.201	0.028	−2.684	0.007	−1.41	0.159
	Submucosal volume (mm3)	0.515 (0.457, 0.729)	1.389 (1.371, 2.588)	1.484 (0.972, 1.795)	−3.175	<0.001	−3.123	0.002	−0.94	0.347
	Muscular volume (mm3)	1.553 (1.054, 2.335)	2.705 (1.992, 3.137)	2.380 (1.863, 3.308)	−2.074	0.039	−2.098	0.036	−0.651	0.515
	Intestinal wall volume (mm3)	5.083 (3.862, 6.203)	9.063 (7.092, 9.934)	7.861 (6.601, 8.674)	−2.921	0.003	−3.025	0.002	−1.229	0.219
	Submucosal volume ratio (%)	12.8 (10.383, 15.327)	16.399 (15.088, 21.089)	17.862 (12.495, 22.586)	−2.244	0.025	−2.586	0.01	−0.217	0.828

All data (nonnormally distributed) are presented as the median (25th, 75th percentiles). Between-group comparisons were performed via the Mann-Whitney U test (nonparametric), and P<0.05 was considered statistically significant. OCT analysis was based on the subset (n=21) with matched three-dimensional pathological data (8 healthy controls, 6 UC mice, and 7 CD mice). CD, Crohn disease; OCT, optical coherence tomography; UC, ulcerative colitis.

Image texture analysis revealed disruption of tissue microstructure

Image texture analysis of both OCT and pathological images was performed to investigate the alterations in tissue microstructural organization (Table 3). OCT entropy was significantly higher in both disease groups as compared with the control group (control: 5.945±0.337; UC: 6.259±0.346, P=0.002; CD: 6.413±0.285, P<0.001), and OCT energy was correspondingly lower (control: 0.321±0.049; UC: 0.289±0.046, P=0.018; CD: 0.272±0.043, P<0.001). The mean OCT gray-scale intensity was significantly reduced in both disease groups (control: 119.840, IQR, 109.030–121.450; UC: 100.480, IQR, 89.665–105.470, P<0.001; CD: 100.565, IQR, 95.750–103.920, P<0.001). In pathological image analysis, contrast differed significantly between disease groups (control: 70.495, IQR, 58.780–91.860; UC: 62.160, IQR, 53.390–69.275, P=0.021; CD: 153.425, IQR, 134.020–184.560, P<0.001). OCT texture parameters differed significantly between the disease groups and the control group (P<0.05 for all comparisons; Table 3). A comprehensive comparison of all parameters is provided in Table S9.

Table 3. Comparative analysis of OCT and pathological image texture parameters between the UC, CD, and control groups.

Category	Parameter	Control group (n=18)	UC group (n=47)	CD group (n=38)	Control vs. UC		Control vs. CD		UC vs. CD
					Z/T	P	Z/T	P	Z/T	P
OCT image	Gray mean	119.840 (109.030, 121.450)	100.480 (89.665, 105.470)	100.565 (95.750, 103.920)	−3.738	<0.001	−3.596	<0.001	−0.557	0.578
	Energy	0.321±0.049	0.289±0.046	0.272±0.043	2.435	0.018	3.769	<0.001	1.777	0.079
	Entropy	5.945±0.337	6.259±0.346	6.413±0.285	−3.286	0.002	−5.399	<0.001	−2.206	0.030
Pathological image	Contrast	70.495 (58.780, 91.860)	62.160 (53.390, 69.275)	153.425 (134.020, 184.560)	−2.302	0.021	−5.509	<0.001	−7.478	0.001

Data following a normal distribution are presented as the mean ± SD (analyzed via the t test); nonnormally distributed data are presented as the median (25th, 75th percentiles) (analyzed via the Mann-Whitney U test). P<0.05 indicates a statistically significant difference. CD, Crohn disease; OCT, optical coherence tomography; SD, standard deviation; UC, ulcerative colitis.

Gray-scale histogram analysis indicated altered optical properties

To assess the changes in overall tissue optical properties of OCT images and in the staining intensity of histopathological images, gray-scale histogram analysis was conducted (Table 4).

Table 4. Comparative analysis of the pathological and OCT image texture parameters and gray-scale distribution ratio between the UC, CD, and control groups.

Category	Parameter	Control group (n=18)	UC group (n=47)	CD group (n=38)	Control vs. UC		Control vs. CD		UC vs. CD
					T/Z	P	T/Z	P	T/Z	P
OCT image	Gray mean	119.840 (109.030, 121.450)	100.480 (89.665, 105.470)	100.565 (95.750, 103.920)	−3.738	<0.001	−3.596	<0.001	−0.557	0.578
	Overall 0 to 49-gray-scale ratio	5.799 (4.783, 15.431)	21.827 (15.936, 26.182)	22.762 (19.285, 27.773)	−4.339	<0.001	−4.649	<0.001	−0.849	0.396
	Overall 100 to 149-gray-scale ratio	43.777 (33.396, 46.665)	31.070 (25.862, 33.808)	29.129 (24.661, 32.967)	−4.339	<0.001	−4.649	<0.001	−1.352	0.176
	Overall 150 to 199-gray-scale ratio	22.088 (18.685, 23.965)	17.201 (12.683, 20.994)	19.361 (16.892, 21.322)	−2.522	0.012	−2.211	0.027	−1.625	0.104
Pathological images	Muscularis 0 to 50-gray-scale ratio (%)	0.002 (0, 0.008)	0.062 (0.047, 0.087)	0.049 (0.043, 0.077)	−4.624	<0.001	−3.557	<0.001	−2.554	0.011

All data (nonnormally distributed) are presented as the median (25th, 75th percentiles). Between-group comparisons were performed via the Mann-Whitney U test (nonparametric), and P<0.05 was considered statistically significant. CD, Crohn disease; OCT, optical coherence tomography; UC, ulcerative colitis.

OCT analysis indicated a decrease in mean gray-scale intensity in both the UC (100.480, IQR, 89.665–105.470) and CD (100.565, IQR, 95.750–103.920) groups compared with the control group (119.840, IQR, 109.030–121.450) (both P values <0.001). Moreover, the disease groups had higher of proportion of pixels in the 0 to 49-gray-scale range (control: 5.799%, IQR, 4.783–15.431%; UC: 21.827%, IQR, 15.936–26.182%, P<0.001; CD: 22.762%, IQR, 19.285–27.773%, P<0.001) but a lower proportion in the 100–149 range (control: 43.777%, IQR, 33.396–46.665%; UC: 31.070%, IQR, 25.862–33.808%, P<0.001; CD: 29.129%, IQR, 24.661–32.967%, P<0.001) and 150–199 range (control: 22.088%, IQR, 18.685–23.965%; UC: 17.201%, IQR, 12.683–20.994%, P=0.012; CD: 19.361%, IQR, 16.892–21.322%, P=0.027). In pathological images, the proportion of pixels in the 0 to 50-gray-scale range within the muscularis layer was increased in both disease groups (control: 0.002%, IQR, 0–0.008%; UC: 0.062%, IQR, 0.047–0.087%, P<0.001; CD: 0.049%, IQR, 0.043–0.077%, P<0.001). Key results from the gray-scale histogram analysis are presented in Table 4, while a detailed comparison of all parameters is available in Table S7.

The integrated multimodal model achieved superior diagnostic performance

ROC curve analysis was performed to evaluate the diagnostic performance of individual parameters and their integration (Figure 5). Individual parameters from the analytical dimensions had AUC values ranging from 0.96 to 0.98 (Figures S4-S7). For example, the OCT-derived mucosal fold angle yielded an AUC of 0.978, pathological image contrast an AUC of 0.963, and the OCT-derived overall 0 to 49-gray-scale ratio an AUC of 0.969.

Figure 5.

ROC curve analysis. (A) ROC curves for UC planar parameters. (B) ROC curves for UC three-dimensional stereoscopic parameters. (C) ROC curves for UC image texture parameters. (D) ROC curves for the UC image gray-scale histogram. (E) ROC curves for CD planar parameters. (F) ROC curves for CD three-dimensional stereoscopic parameters. (G) ROC curves for CD image texture parameters. (H) ROC curves for CD image gray-scale histogram. (I) ROC curve analysis integrating all quantitative parameters from UC imaging. (J) ROC curve analysis integrating all quantitative parameters from CD imaging. ROC curve analysis integrating all quantitative parameters for UC and CD imaging, including planar measurements, three-dimensional volumetric data, image texture features, and gray-scale histogram characteristics. AUC, area under the curve; CD, Crohn disease; OCT, optical coherence tomography; ROC, receiver operating characteristic; UC, ulcerative colitis.

Feature integration was performed across planar, texture, and gray-scale analyses from both pathology and OCT modalities. Three-dimensional volumetric parameters, available for only 21 matched samples, were excluded from the primary predictive model to reduce overfitting risk and were used separately for independent confirmatory analysis. From an initial pool of 84 two-dimensional features, a two-stage feature selection procedure (univariate filtering followed by recursive feature elimination with random forest) embedded within LOOCV identified 19 predictors, and all preprocessing and selection steps were performed within each training fold to prevent data leakage. According to LOOCV, the integrated model achieved an AUC of 1.000 for discriminating UC and CD mice from control mice (Figure 5).

Discussion

The clinical assessment of intestinal inflammation remains challenging, particularly as transmural healing for patients with CD has emerged as both a critical therapeutic objective and a novel treatment target (35). Current evaluation frameworks for transmural healing incorporate multiple dimensions, including endoscopic, histopathological, and molecular level assessments (36,37). The pursuit of deeper healing end points, such as transmural healing, therefore requires technologies capable of providing comprehensive evaluations across multiple biological scales. At present, clinical practice relies heavily on cross-sectional radiological imaging modalities, including magnetic resonance enterography and computed tomography enterography, to visualize inflammatory changes across the full thickness of the bowel wall (38).

This study established a novel quantitative histopathological system for murine IBD by extracting multidimensional features from pathological sections, including 2D morphology, 3D architecture, and textural patterns. Building on this framework, we further developed a co-registered OCT imaging system that leverages its optical biopsy capability and, through the application of endoscopic OCT, enables noninvasive, real-time, and quantitative assessment of intestinal inflammation. In our preliminary investigations, we discovered that key OCT-derived parameters, specifically the mucosal layer angle and the combined thickness of the submucosa and muscularis propria, possess significant value for stratifying inflammatory severity in murine models (39). Additionally, these preliminary studies further highlighted the important role of OCT in differentiating between colorectal dysplasia and cancer and between the acute and chronic forms of colitis in mice (40,41). Building upon these foundational findings, we sought to identify novel imaging biomarkers to further enhance diagnostic precision. For the first time, our results demonstrated that a synergistic multimodal and multidimensional analytical strategy not only enables robust quantification of IBD-related changes but also provides superior diagnostic performance as compared with any single analytical approach. A suite of powerful biomarkers was successfully identified across planar, volumetric, textural, and gray-scale domains.

A central finding of this work is the pronounced synergistic effect observed through the integration of multimodal data. Although the integrated model yielded a perfect average AUC in our limited dataset according to LOOCV—with feature reduction applied prior to model training to mitigate overfitting—this outcome should be interpreted primarily as a strong indicator of informational synergy across modalities rather than as proof of generalizable diagnostic accuracy in external cohorts. The risk of model overfitting is substantial, and the perfect score should be interpreted as a ceiling effect under idealized conditions. The true scientific value lies not in the absolute metric itself but rather in the marked improvement in the AUC, from 0.987 for the high-performing single-modality features to a perfect score following integration. This observation strongly supports our central hypothesis that the complexity of IBD pathology cannot be fully captured through any single analytical lens and that a holistic, multifaceted approach is necessary for generating a more complete digital signature of the disease (42).

Planar model analysis revealed that both the CD and UC groups exhibited significant increases in mucosal fold angle, mucosal fold thickness, submucosal area, and submucosal cell count as compared with controls in both the histological and OCT assessments, suggesting that inflammatory edema involves the mucosal and submucosal layers (18,43). These pathological alterations result in the flattening of mucosal folds and consequent enlargement of fold angles, which were effectively captured with quantitative 2D morphometric parameters (44,45).

To address the inherent sampling limitations of 2D histology, a 3D analytical strategy was employed to provide a more comprehensive and biologically meaningful representation of inflammation (46). As previously reported (47), 3D-OCT enables direct volumetric acquisition, whereas pathological 3D space can be reconstructed through the stacking of sequential 2D sections, a concept extended in this study through the proposed bucket model. This model indicates that the subtle variations detected in 2D morphometry are amplified when accumulated across tissue volumes, thereby enhancing sensitivity for inflammation detection. In the analyses of the murine UC and CD models, significant increases were observed in submucosal cell count, mucosal volume, submucosal volume, and total intestinal wall volume, accompanied by a reduction in gland count (AUC =0.679–1.000; 95% confidence interval: 0.378–1.000) (Figure 5). These findings suggest that the expansion of inflammatory tissue contributes to the progressive loss of functional glandular architecture, which reflects disease severity. The intrinsic 3D volumetric acquisition capability of OCT facilitated efficient calculation of layer-specific volumes and spatial relationships (48,49), which were subsequently validated through pathological analysis (Figure 5).

Beyond structural parameters, this framework introduces novel biomarkers derived from image texture and gray-scale analyses. Texture analysis demonstrated decreased energy and increased entropy in inflamed tissues, quantitatively reflecting loss of structural homogeneity. Notably, the divergence in contrast between UC models, which exhibited lower contrast, and CD models, which exhibited higher contrast, suggests that these textural features may serve as noninvasive indicators reflecting distinct pathological patterns, specifically diffuse versus focal inflammation. The ability of this phenomenon to serve as a tool for differentiating IBD subtypes should be further investigated. The characteristic leftward shift observed in the OCT gray-scale histograms represents a direct physical manifestation of increased light scattering in inflamed tissue rather than a purely statistical phenomenon. Accordingly, the proposed framework effectively translates complex biophysical changes into objective and computable imaging metrics (50).

The improvement from individual predictors (AUC range 0.96–0.98) to the integrated model (AUC =1.000) suggests a synergistic contribution of multimodal and multidimensional data integration. This informational synergy, rather than the absolute AUC value, merits emphasis, given the limited sample size and the risk of overfitting.

The gray-scale histogram changes (Table 4) indicated a significant increase in low-gray-scale pixels (0–49 range) in diseased tissues (control: 5.799%, IQR, 4.783–15.431%; UC: 21.827%, IQR, 15.936–26.182%; CD: 22.762%, IQR, 19.285–27.773%; all P values <0.001), consistent with the increased light scattering and signal attenuation arising secondary to inflammatory cell infiltration and tissue remodeling.

The texture analysis (Table 3) revealed a loss of tissue homogeneity and structural regularity in inflamed tissues, characterized by increased entropy (UC: 6.259±0.346; CD: 6.413±0.285; control: 5.945±0.337) and decreased energy (UC: 0.289±0.046; CD: 0.272±0.043; control: 0.321±0.049), reflecting disruption of the organized and repetitive microarchitecture of healthy tissue.

The 3D volumetric data suggest that the glandular and cellular changes (summarized in Table 2) underlie the volumetric alterations observed in both modalities. The agreement between OCT and histopathological volumetry (ρ=0.487–0.944; all P values <0.05; Figure S6) supports OCT as a surrogate modality for 3D histopathological quantification, with the systematic positive bias (mean difference 0.75–4.87 mm³) consistent with the additional scattering signal captured by OCT imaging.

The concordance between histopathological and OCT planar findings, particularly for mucosal fold angle and layer-specific areas (Table 1), supports the potential of OCT as a complementary modality for the noninvasive assessment of IBD-related structural changes. The mucosal fold angle from pathological images showed a correlation with overall disease severity (r=0.755; P<0.001), suggesting its utility as an indicator of inflammatory severity.

Conclusions

This study successfully established and validated two parallel diagnostic systems by integrating digital pathology and OCT-based tissue imaging, providing a novel avenue for the precision assessment of IBD. Through systematic multimodal and multidimensional analyses consisting of 2D planar, 3D volumetric, textural, and gray-scale features, this work achieved, for the first time, a comprehensive digital characterization of IBD-related pathological alterations and demonstrated that the integrated diagnostic model provides superior performance compared with any single method. Specifically, 2D planar parameters effectively quantified the tissue remodeling associated with inflammatory edema, 3D volumetric features captured the cumulative impact of subtle morphological changes through the bucket model. Moreover, texture analysis objectively reflected the loss of tissue homogeneity through distinct energy entropy patterns, with contrast differences offering new insights into distinguishing UC and CD from healthy tissue. The characteristic leftward shift of OCT gray-scale histograms may facilitate the translation of optical scattering properties into quantifiable metrics, and the marked performance improvement of the integrated model, with the AUC increasing from 0.987 to 1.000, underscores the necessity of multidimensional feature fusion for capturing the complexity of IBD pathology.

The principal strength of this framework lies in data integration, which, in our study, generated distinct quantitative signatures corresponding to UC-like diffuse and homogeneous inflammation and CD-like transmural and high contrast inflammation, constituting a novel strategy for the noninvasive subtyping of IBD (51,52). This work both provides a set of quantifiable imaging biomarkers and establishes a methodological foundation for noninvasive precision diagnostics. Future studies will focus on improving the model’s generalizability in larger cohorts and exploring its potential applications in disease staging and the prediction of therapeutic response.

Supplementary

The article’s supplementary files as

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. All animal procedures were approved by the Institutional Animal Care and Use Committee of Fujian Medical University (approval No. IACUCFJMU2024-Y-2494), and were conducted in accordance with institutional guidelines for the care and use of laboratory animals.

Data Sharing Statement

Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0232/dss

qims-16-10-772-dss.pdf ^{(115.4KB, pdf)}

DOI: 10.21037/qims-2026-0232