Development of a Computed Tomography Enterography-Based Radiomics Model to Characterize Inflammation, Fibrosis, and Smooth Muscle Thickening in Stricturing Crohn’s Disease

Abstract

INTRODUCTION:

Computed tomography enterography (CTE) is a noninvasive cross-sectional imaging modality routinely used for diagnosis of Crohn’s disease (CD) strictures. However, CTE is limited in individually defining the severity of inflammation, fibrosis, or smooth muscle hyperplasia/hypertrophy (SMH) within such strictures. We developed and internally validated a radiomics-based machine learning model for separately characterizing the degree of histopathological inflammation, fibrosis, and SMH in CD strictures and compared it with centrally read visual radiologist scoring of CTE.

METHODS:

This single-center, cross-sectional study included 100 patients with CD (n = 66 for discovery; n = 34 for validation) with terminal ileal strictures confirmed on diagnostic CTE within 15 weeks of intestinal resection. Histopathological specimens were scored for inflammation, fibrosis, and SMH and spatially linked with corresponding pre-surgical CTE sequences. Annotated stricture regions on CTE were scored visually by radiologists as well as underwent three-dimensional radiomics-based machine-learning analysis; both evaluated against histopathology scoring.

RESULTS:

Three distinct sets of radiomic features capturing textural heterogeneity within strictures were separately linked with severe inflammation, fibrosis, and SMH across both discovery (accuracy = 0.74, 0.71, 0.71, respectively) and validation (accuracy = 0.74, 0.68, 0.68, respectively) cohorts. Radiologist visual scoring was limited, with accuracy of 0.51 for identifying severe inflammation, 0.48 for severe fibrosis, and 0.58 for severe SMH. Augmenting radiologist scoring with radiomic features yielded improved performance in identifying severe inflammation, fibrosis, or SMH compared to radiologist scoring alone.

DISCUSSION:

Radiomic features of CD strictures on CTE can identify severe histopathological fibrosis, inflammation, and smooth muscle hyperplasia and perform better than radiologist visual scoring in stricture characterization.

Graphical Abstract

INTRODUCTION

Crohn’s disease (CD) strictures present a clinical challenge to gastroenterologists, owing to their limited response to advanced anti-inflammatory advanced therapies resulting in a need for surgical intervention. Both smooth muscle cell hyperplasia and hypertrophy (SMH) as well as non-inflammatory pathways of fibrosis are pivotal to the progression of fibrostenotic strictures, even when the inflammatory component is addressed (1,2). Efforts are thus underway to develop therapeutic targets for all 3 of these stricture components (1). The gold standard for subtyping of stricturing CD thus remains surgical resection, based on which a detailed histopathological assessment of inflammation, fibrosis, and SMH can be obtained (3). However, the clinical workflow relies on advanced imaging modalities shown to accurately detect strictures, which are limited in assessment of the level of fibrosis, SMH, and inflammation separately within a given stricture (4).

Radiomics is a machine learning (ML)-based method that extracts quantitative features from medical images using data characterization algorithms (5). Previous studies have evaluated radiomics models to characterize CD strictures, on both magnetic resonance enterography and computed tomography enterography (CTE), yet none have attempted to characterize SMH separately from inflammatory or fibrostenotic components (6,7). This distinction has crucial clinical importance as SMH is a key contributor to luminal narrowing in CD strictures.

In this study, we leveraged a carefully curated cohort of CTE scans with pathologically linked and scored surgical specimens of CD strictures to develop and validate radiomics models for characterizing pathological inflammation, fibrosis, and SMH. We hypothesized that radiomic features could be used to distinguish histopathologically severe from less severe inflammatory, fibrotic, or muscular CD strictures, either alone or when combined with centrally read and standardized radiologist assessment.

METHODS

Patient population

For this retrospective, cross-sectional, quaternary single-center minimal risk research study, electronic medical records were queried for adult patients with CD from January 2007 to March 2021. Inclusion criteria were as follows: (i) men or nonpregnant women, age ≥18 years; (ii) guideline-based diagnosis of CD (8); (iii) symptomatic nonmalignant CD stricture on CTE (meeting the CONSTRICT criteria (9)) within 15 cm of the ileocecal valve (for naïve) or the anastomosis (for anastomotic strictures); (iv) CTE performed ≤15 weeks before surgical resection; and (v) surgical pathology slides with accurate spatial match to the area of maximal stricture narrowing available. Exclusion criteria were the following: (i) stricture with complications such as abscess, phlegmon (a history of penetrating disease treated with resection is allowed if not present at the time of assessment); (ii) stricturoplasty in the distal ileum within 1 year of imaging; (iii) terminal ileal stricture in subject with end ileostomy; (iv) diverting loop ileostomy proximal to the dominant stricture; and (v) ileoanal or Kock pouch. Additional detailed criteria are in Supplemental Table 1 (http://links.lww.com/AJG/D952). Imaging and surgical histopathology were scored and/or annotated by an inflammatory bowel disease (IBD) radiologist (N.S.G.) and IBD pathologist (I.O.G.), respectively (both with more than 15 years of experience). Patient demographics and clinical characteristics were confirmed to ensure compliance with inclusion/exclusion criteria and stored in a secure online database for later analysis.

Histopathological grading

Histopathological scoring was performed on all specimens, in accordance with an international consensus standard (10), using a pathologist’s Visual Analog Scale (pVAS; range 0 for normal to 100 for the most severe degree ever encountered) to provide an overall histopathological assessment for severity of inflammation, fibrosis, and SMH, separately. Based on empirical evaluation of the cohort pVAS score distributions and their medians, thresholds were selected to dichotomize each pVAS score while ensuring an approximately even split of severe and nonsevere patients for each subtype. These thresholds (inflammation pVAS ≥ 60, fibrosis pVAS ≥ 40, SMH pVAS ≥ 50) were used to define severe phenotypes for ML model construction by reflecting relative severity within this surgical cohort, in the absence of universally accepted thresholds for clinical severity. Supplementary Figure 1 (http://links.lww.com/AJG/D952) depicts the distributions for pVAS scores corresponding to inflammation, SMH, and fibrosis in our cohort.

CTE image annotation and radiologist assessment

All CTE sequences were used to identify terminal/neoterminal ileal strictures based on the CONSTRICT criteria (9). CTE acquisition involved small bowel distension with 1 L of neutral or low-density oral contrast mixture and abdominal/pelvic CT examination during the enteric phase after the administration of intravenous contrast (see Supplementary Methods [http://links.lww.com/AJG/D952] for details of radiology image acquisition and standardized scoring). Before annotation, the study pathologist and radiologist conferred to ensure that regions of interest (ROIs) were in spatial correspondence between histopathology and CTE by identifying the most narrowed stricture area on both modalities where bowel wall thickening was present. Only a single dominant stricture ROI (in-plane thickness of 3 mm centered on the bowel wall) located within 15 cm of the ileocecal valve or anastomosis was annotated and scored per CTE, which encompassed the full length of the dominant stricture in 3D. This volumetric annotation strategy was chosen to capture spatial heterogeneity along the entire stricture while reducing sensitivity to local bowel curvature or out-of-plane orientation. One similar thickness nonstricture ROI was additionally and separately placed on the adjacent healthy small bowel per patient. A radiological VAS (rVAS) was recorded for severity of inflammation (range: 0 to 100, with 100 being the most severe inflammation ever encountered; “inflammation rVAS”). Given that both muscular and fibrotic changes affect the assessment of non-inflammatory changes on CTE, an rVAS was recorded for the severity of noninflammatory changes to reflect both fibrosis and SMH (range: 0 to 100, with 100 being the most severe non-inflammatory changes ever encountered; “fibrostenosis rVAS”). CTE and histopathology scoring were performed in a blinded fashion and independently from each other.

Postprocessing of CTE images and radiomic feature extraction

Figure 1 summarizes overall experimental workflow for radiomic analysis. A publicly available radiology imaging quality control package (https://viswanath-lab.github.io/RadQy/) (11) was used to determine and correct any imaging artifacts in CTE sequences selected for analysis. Radiomic features were extracted on a pervoxel basis within annotated 3D stricture ROIs and 3D healthy ROIs separately. Purely geometric or shape-based radiomic features (e.g., thickness, elongation, surface area) were excluded from model development because of their susceptibility to distortion in regions of out-of-plane bowel angulation. To assess whether regions proximal to the annotated portion of the bowel wall could further aid in stricture characterization, morphological dilation was applied to the annotated ROI to obtain 3 mm and 6 mm expansions (of the original stricture and healthy annotations) from which radiomic features were then extracted. This allowed the capture of subtle variations in image appearance and heterogeneity within as well as around the strictured bowel wall on CTE scans based on responses to different texture operators. Voxel-level feature responses were subsequently aggregated using region-level statistical summaries (e.g., mean and variance) and associated with the original annotated ROIs as well as the 3 and 6 mm bandexpansions. Supplementary Table 2 (http://links.lww.com/AJG/D952) summarizes Image Biomarker Standardization Initiative-compliant definitions for all radiomic features used in this study.

Figure 1.

Experimental workflow for developing the stricturing CD radiomics model through CTE. (a) An IBD radiologist identified and annotated the region of TI stricturing disease in conjunction with an IBD pathologist with access to the resection. Band expansions beyond original TI annotation are performed automatically at 3 and 6 mm diameters. (b) Quantitative radiomic features are extracted on a per-pixel basis within the green regions of interest ROI for each of the TI, 3 and 6 mm bands. (c) Machine learning to select subset of radiomic features that strongly discriminate between pathologically severe and nonsevere inflammation (or fibrosis/SMH), while also being highly reproducible across both stricture and healthy ROIs. (d) Standardized scoring of corresponding pathology section used as ground truth for severity of inflammation, fibrosis, and SMH. (e) Radiomic and histopathology scores used to train a ML model. (f) ML model evaluated for predicting severity of histopathological findings within the stricturing region based on the radiomic profile. CTE, computed tomography enterography; IBD, inflammatory bowel disease; ML, machine learning; pVAS, pathologist’s Visual Analog Scale; ROIs, regions of interest; SMH, smooth muscle hypertrophy/hyperplasia; TI, terminal ileum.

Statistical analysis

All included patient studies were first segregated into discovery (comprising a randomly selected set of 66% of patients considered) and hold-out validation cohorts (comprising the remaining patients). Discovery and validation cohorts were separately constructed based on patient groupings for each of (i) pathologically severe/nonsevere inflammation, (ii) pathologically severe/nonsevere fibrosis, and (iii) pathologically severe/nonsevere SMH. Categorical clinical variables were described as proportions or percentages and statistically compared between discovery and validation subcohorts through the Fisher exact test. Continuous variables including age at diagnosis of CD, median time duration between CD diagnosis and stricture diagnosis, median duration from stricture diagnosis to surgical resection, and medical, endoscopic, and surgical treatments were summarized using means and ranges using the Mann-Whitney U test. The criterion for statistical significance was P < 0.05.

Each of the inflammation and fibrostenotic rVAS scores were separately thresholded at fixed intervals to evaluate their performance in distinguishing between each of the pathologically severe/nonsevere subtypes (inflammation, fibrosis, SMH), respectively, quantified through area under the receiver operator curve (AUROC). Corresponding accuracy, sensitivity, specificity, and F1 scores for each of the inflammation and fibrostenotic rVAS scores were also calculated within the validation subcohort, based on thresholding rVAS scores at 50 (assumed to correspond to equivocal radiological evaluation).

Three radiomics models were constructed to independently predict pathologically severe inflammation, severe SMH, and severe fibrosis by identifying the most relevant and reproducible radiomic descriptors associated with original ROIs for each subtype (severe/nonsevere inflammation, severe/nonsevere SMH, severe/nonsevere fibrosis) separately, through the ML approach described in the Supplementary Materials (http://links.lww.com/AJG/D952). The performance of each radiomic model was evaluated through AUROC of associated random forest (RF) classifier models (12), calculated within corresponding discovery and hold-out validation sub-cohorts for each of severe/nonsevere inflammation, severe/nonsevere SMH, and severe/nonsevere fibrosis. Corresponding accuracy, sensitivity, specificity, and F1 scores for each of the inflammation, SMH, and fibrosis radiomic models were additionally calculated based on the optimal threshold identified through each discovery cohort.

To assess whether radiological evaluation augmented radiomic model performance, top-ranked radiomic features associated with each of severe inflammation, severe fibrosis, and severe SMH were paired with the corresponding inflammation and fibrostenosis rVAS scores (where the latter was used in conjunction with both SMH and fibrosis radiomic features). These augmented radiomics/rVAS feature sets were used to construct separate RF models for each subtype, with AUROC used as a measure of classifier performance. As radiomics and rVAS scores were collected at separate times, assessment of the impact of the radiomics model on real-time radiologist scoring was not conducted. The DeLong test was used to assess significant differences in AUROC performance for identifying pathologically severe inflammation, severe fibrosis, or severe SMH within corresponding hold-out validation cohorts, through pairwise comparisons between the radiomic model, rVAS score, and the augmented radiomics/rVAS feature set (conducted for each subtype separately).

Ethical statement

This study was approved by the Institutional Review Board at the Cleveland Clinic under IRB number 20–916.

RESULTS

Cohort demographics

One hundred patients with CD met inclusion criteria. Demographics were reflective of a quaternary referral center cohort of CD patients with stricturing disease. The median age was 39 years (range 18–81 years); 50% were female and 87% White (Table 1). Per Montreal classification, 82% of the patients solely had stricturing disease (B2) whereas 18% had a history of internal penetrating disease (B3). Perianal disease was present in 28% of the cohort at any time in the disease. Eleven percent of patients had 2 identifiable ileal strictures; 53% of the strictures were anastomotic, and 13% had endoscopic balloon dilation within a median duration time of 2 months (range 0.3–30 months) before imaging. Only one patient had a previous endoscopic stricturotomy 22 months before the surgical resection. The median duration between CD diagnosis and stricture diagnosis was 13 years (range 0–72 years) and median duration from stricture diagnosis to surgical resection was 10 months (range 0–108 months), respectively. CTE was performed at a median of 4 weeks (range 0.1–14.4 weeks) before surgical resection of the strictures. Seventy-eight percent of patients were exposed to biologics and immunomodulators within 8 weeks of imaging. At the time of surgery, 84% of patients had symptoms consistent with obstruction with nausea/vomiting (46%) and abdominal pain (73%) being the most common. Statistical evaluation demonstrated no significant differences in clinical, demographic, and treatment information between discovery and validation cohorts for each pathological subtype (severe/nonsevere inflammation, severe/nonsevere SMH, severe/nonsevere fibrosis).

Table 1.

Demographics and baseline clinical features of study population summarizing no significant differences between discovery and validation cohorts constructed for pathological fibrosis, inflammation, and SMH subtypes

			Grouping based on pathological inflammation scoring			Grouping based on pathological fibrosis scoring			Grouping based on pathological SMH scoring
Category/variable	Groups/statistics	All N = 100	Discovery cohort (N = 66)	Validation cohort (N = 34)	P value	Discovery cohort (N = 66)	Validation cohort (N = 34)	P value	Discovery cohort (N = 66)	Validation cohort (N = 34)	P value
Sex	Male n (%)	50 (50)	30	20	0.21	32	18	0.67	27	23	0.01
	Female n (%)	50 (50)	36	14		34	16		39	11
Age at CD diagnosis, yr	Median (range)	21 (2–75)	21 (2–62)	21 (6–75)	0.5	21 (2–62)	21 (7–75)	0.33	22 (6–75)	18.5 (2–56)	0.12
Age at stricture dx, yr	Median (range)	21 (2–75)	21 (2–62)	21 (6–75)	0.5	21 (2–62)	21 (7–75)	0.33	22 (6–75)	18.5 (2–56)	0.12
Age at MRE, yr	Median (range)	39 (18–81)	41 (18–75)	37.5 (20–81)	0.47	42 (19–77)	35 (18–81)	0.17	39.5 (20–81)	37.5 (18–75)	0.88
Duration between CD & stricture diagnosis, yr	Median (range)	13 (0–72)	12.5 (1–42)	13.5 (1–42)	0.42	14.5 (0–72)	8 (0–39)	0.015	13 (0–49)	13 (0–72)	0.53
Duration between stricture diagnosis and surgery, mo	Median (range)	10 (0–108)	8.5 (0–108)	10.5 (0–53)	0.81	10.5 (0–100)	8.5 (0–108)	0.73	9.5 (0–100)	11 (0–108)	0.77
Duration between MRE and resection, wk	Median (range)	4 (0.1–14.4)	4.9 (0.1–14.3)	3.3 (0.1–14.4)	0.27	4.2 (0.1–14.4)	3.7 (0.1–14.3)	0.23	4.2 (0.1–14.4)	3.1 (0.1–14.4)	0.97
Obstructive symptoms at time of imaging	n (%)	84 (84.0)	58 (87.9)	26 (76.5)	0.16	54 (81.8)	30 (88.2)	0.57	58 (87.9)	26 (76.5)	0.16
History of extraintestinal manifestationa	n (%)	64 (64.0)	41 (62.1)	23 (67.6)	0.59	43 (65.2)	21 (61.8)	0.74	41 (62.1)	23 (67.6)	0.59
Ileocecal resection prior to current stricture	n (%)	54 (54.0)	38 (57.6)	16 (47.1)	0.40	38 (57.6)	16 (47.1)	0.40	37 (56.1)	17 (50.0)	0.67
Number of resections in those with prior resections	Median (range)	2 (1–8)	2 (1–8)	1.5 (1–8)	0.73	2 (1–8)	1.5 (1–4)	0.80	1 (1–8)	2 (1–4)	0.53
CD Montreal classification (at any time in the disease course)	B2 (Stricturing) n (%)	61 (61.0)	41 (62.1)	20 (58.8)	0.75	43 (65.2)	18 (52.9)	0.22	37 (56.1)	24 (70.6)	0.52
	B2p (Stricturing with perianal disease) n (%)	21 (21.0)	15 (22.7)	6 (17.6)		15 (22.7)	6 (17.6)		15 (22.7)	6 (17.6)
	B3 (Fistulizing) n (%)	11 (11.0)	6 (9.1)	5 (14.7)		5 (7.6)	6 (17.6)		9 (13.6)	2 (5.9)
	B3p (Fistulizing with perianal disease) n (%)	7 (7.0)	4 (6.1)	3 (8.8)		3 (4.5)	4 (11.8)		5 (7.6)	2 (5.9)
Type of stricture	Anastomotic n (%)	53 (53.0)	37 (56.1)	16 (47.1)	0.41	37 (56.1)	16 (47.1)	0.407	36 (54.5)	17 (50.0)	0.68
	Naive n (%)	47 (47.0)	29 (43.9)	18 (52.9)		29 (43.9)	18 (52.9)		30 (45.5)	17 (50.0)
Medications for CD within 8 wk from imaging	5-aminosalicylic-acid, oral or rectal n (%)	17 (17.0)	7 (10.6)	10 (29.4)	0.03	15 (22.7)	2 (5.9)	0.048	11 (16.7)	6 (17.6)	1.00
	Steroid, systematic, n (%)	41 (41.0)	27 (40.9)	14 (41.2)	1	24 (36.4)	17 (50.0)	0.205	31 (47.0)	10 (29.4)	0.13
	Steroid, rectal or Budesonide, n (%)	14 (14.0)	11 (16.7)	3 (8.8)	0.37	8 (12.1)	6 (17.6)	0.545	9 (13.6)	5 (14.7)	1
	Mercaptopurine or Azathioprine n (%)	14 (14.0)	10 (15.2)	4 (11.8)	0.77	12 (18.2)	2 (5.9)	0.13	6 (9.1)	8 (23.5)	0.07
	Methotrexate n (%)	7 (7.0)	4 (6.1)	3 (8.8)	0.69	3 (4.5)	4 (11.8)	0.22	4 (6.1)	3 (8.8)	0.69
Medications for CD within 8 wk from imaging	Certolizumab n (%)	2 (2.0)	1 (1.5)	1 (2.9)	1.00	2 (3.0)	0 (0.0)	0.55	2 (3.0)	0 (0.0)	0.55
	Adalimumab n (%)	16 (16.0)	10 (15.2)	6 (17.6)	0.78	8 (12.1)	8 (23.5)	0.16	10 (15.2)	6 (17.6)	0.79
	Infliximab n (%)	12 (12.0)	10 (15.2)	2 (5.9)	0.21	7 (10.6)	5 (14.7)	0.54	6 (9.1)	6 (17.6)	0.33
	Vedolizumab n (%)	4 (4.0)	3 (4.5)	1 (2.9)	1.00	4 (6.1)	0 (0.0)	0.30	2 (3.0)	2 (5.9)	0.60
	Ustekinumab n (%)	7 (7.0)	5 (7.6)	2 (5.9)	1.00	4 (6.1)	3 (8.8)	0.69	5 (7.6)	2 (5.9)	1.00
	None n (%)	22 (22.0)	15 (22.7)	7 (20.6)	1.00	17 (25.8)	5 (14.7)	0.31	14 (21.2)	8 (23.5)	0.80
Pathologist Visual Analog Scoring
Severity of inflammation	Median (range)	62 (2–99)	62 (2–99)	61.5 (2–93)	0.50	63 (2–99)	59 (2–94)	0.92	62 (2–95)	61.5 (2–99)	0.95
Severity of fibrosis	Median (range)	42.5 (4–93)	50 (9–93)	37.5 (4–91)	0.33	43.5 (9–93)	41 (13–91)	0.58	42 (4–91)	50 (11–93)	0.54
Severity of SMH	Median (range)	52.5 (6–97)	56 (6–93)	48 (7–97)	0.34	46 (6–97)	57 (6–91)	0.48	51 (6–97)	60.5 (6–92)	0.57

CD, Crohn’s disease CTE, computed tomography enterography; dx, diagnosis; MRE, magnetic resonance elastography; N, number; SMH, smooth muscle hypertrophy/hyperplasia

^aExtraintestinal manifestations included anemia, oral ulcers, hypoalbuminemia and nutritional deficiencies, eye, skin, or joint involvement.

Radiomics analysis can accurately characterize pathologically severe inflammation, fibrosis, and SMH within CD strictures on CTE

Fifteen radiomic descriptors were identified as associated with severe inflammation and comprised texture responses to ripple-edge-level patterns as well as correlation and contrast of gradient co-occurrences (Supplemental Table 3, http://links.lww.com/AJG/D952), with 9 features derived from the bowel wall ROI and 6 features from the 6 mm ROI expansion. Figure 2a–e depict representative heatmap visualizations of radiomic features from annotated stricture ROIs and band expansions, on CTE. Different heterogeneity patterns can be observed within both ROIs in more severely inflamed strictures (more blue/red) vs nonsevere inflammation (primarily red). The RF model based on these radiomic descriptors (Figure 2f-h) accurately discriminated between severe and nonsevere histopathological inflammation with a robust discovery accuracy of 0.74 ± 0.06 (sensitivity = 0.69 ± 0.12, specificity = 0.80 ± 0.11, F1 score = 0.75 ± 0.07, AUROC = 0.73 ± 0.05), and a corresponding hold-out validation accuracy of 0.74 (sensitivity = 0.5, specificity = 1, F1 score = 0.67, AUROC = 0.71). Performance of the individual bowel wall signature (AUROC = 0.70, 0.60) and 6 mm expansion signature (AUROC = 0.65, 0.43) was found to be significantly worse than the integrated radiomic signature (Supplemental Table 4, http://links.lww.com/AJG/D952).

Figure 2.

Visualization of radiomic heatmaps and model performance in discriminating pathologically severe and nonsevere inflammation within stricturing Crohn’s disease through CTE. Depiction of 4 representative patients with nonsevere and severe inflammation from discovery and validation cohorts, respectively, based on histopathology scoring of the resection tissue (a), together with corresponding CTE image (b) with annotated regions of interest for TI (green) and 6 mm band expansion (red) (c) as well as top ranked radiomic feature heatmaps (d and e). Higher heterogeneity in radiomic expression observed through increasing proportion of blue/red within heatmaps of stricture wall annotation and within the 6 mm expansion band. (f) Receiver operating characteristic curves showing ML model performance for top ranked radiomic features (in both discovery and validation cohorts), rVAS score (in the entire cohort), and the augmented rVAS/radiomics model (in the validation cohort). (g) Bar plots for sensitivity, specificity, F1 scores, and accuracy of the models. (h) Average radiomic classifier probability for severe inflammation plotted against ground truth histopathology pVAS. The central solid line shows the linear regression between these 2 variables with the thinner curves showing error bounds. CTE, computed tomography enterography; ML, machine learning; pVAS, pathologist’s Visual Analog Scale; rVAS, radiological VAS; TI, terminal ileum.

Histopathologically severe fibrosis was found to be associated with a different set of 15 radiomic descriptors, comprising primarily intensity and gradient co-occurrence responses from the bowel wall (Supplemental Table 3, http://links.lww.com/AJG/D952) in addition to wavelet/pattern responses from the 3 mm ROI expansion. Representative heatmaps in Figure 3a–e reveal an overall lower expression of Haralick feature responses within severely fibrotic CTE-annotated strictures (more blue) and an overall higher expression (more red) within the 3 mm ROI expansion. By contrast, nonsevere strictures exhibit a mixed expression in both ROIs. The corresponding radiomics-based RF model (Figure 3f-h) yielded a discovery accuracy of 0.71 ± 0.06 (sensitivity = 0.74 ± 0.09, specificity = 0.68 ± 0.08, F1 score = 0.73 ± 0.06, AUROC = 0.70 ± 0.04) in distinguishing between severe and nonsevere pathological fibrosis groups, and a corresponding hold-out validation accuracy of 0.68 (sensitivity = 0.67, specificity = 0.69, F1 score = 0.69, AUROC = 0.66). By comparison, while radiomic features from 3 mm expansion alone yielded similar performance (AUROC = 0.68, 0.60), the bowel wall signature performed markedly worse (AUROC = 0.64, 0.52) (Supplemental Table 4, http://links.lww.com/AJG/D952).

Figure 3.

Visualization of radiomic heatmaps and model performance in discriminating pathologically severe and non-severe fibrosis within stricturing Crohn’s disease through CTE. Depiction of 4 representative patients with nonsevere and severe fibrosis from discovery and validation cohorts, respectively, based on histopathology scoring of the resection tissue (a), together with corresponding CTE image (b) with annotated regions of interest for TI (green) and 3 mm band expansion (red) (c) as well as top ranked radiomic feature heatmaps (d and e). Higher heterogeneity in radiomic expression observed through increasing proportion of blue/red within heatmaps of stricture wall annotation and within the 3 mm expansion band. (f) Receiver operating characteristic curves showing ML model performance for top ranked radiomic features (in both discovery and validation cohorts), rVAS score (in the entire cohort), and the augmented rVAS/radiomics model (in the validation cohort). (g) Bar plots for sensitivity, specificity, F1 scores, and accuracy of the models. (h) Average radiomic classifier probability for severe fibrosis plotted against ground truth histopathology pVAS. The central solid line shows the linear regression between these 2 variables with the thinner curves showing error bounds. CTE, computed tomography enterography; ML, machine learning; pVAS, pathologist’s Visual Analog Scale; rVAS, radiological VAS; TI, terminal ileum.

Radiomics descriptors most associated with histologically severe SMH involved primarily ripple-level patterns from the bowel wall and the 3 mm ROI expansion, in addition to several gradient co-occurrence features (Supplemental Table 3, http://links.lww.com/AJG/D952). Heatmaps shows that both ROIs exhibit similar mixed heterogeneity patterns (green/red/blue) within nonsevere SMH compared with the bowel wall in patients with severe SMH (primarily green, Figure 4a–e). For distinguishing between severe and nonsevere pathological SMH groups, the associated RF machine classifier (Figure 4f-h) yielded a discovery accuracy of 0.71 ± 0.06 (sensitivity = 0.74 ± 0.09, specificity = 0.69 ± 0.08, F1 score = 0.72 ± 0.06, AUROC = 0.63 ± 0.05) and a hold-out validation accuracy of 0.68 (sensitivity = 0.5, specificity = 0.88, F1 score = 0.62, AUROC = 0.63). The individual signatures derived from the bowel wall (state AUROC = 0.58, 0.53) and the 3 mm ROI expansion (AUROC = 0.59, 0.43) were markedly worse than the integration of both regions (Supplemental Table 4, http://links.lww.com/AJG/D952).

Figure 4.

Visualization of radiomic heatmaps and model performance in discriminating pathologically severe and nonsevere SMH within stricturing Crohn’s disease through CTE. Depiction of 4 representative patients with nonsevere and severe SMH from discovery and validation cohorts, respectively, based on histopathology scoring of the resection tissue (a), together with corresponding CTE image (b) with annotated regions of interest for TI (green) and 3 mm band expansion (red) (c) as well as top ranked radiomic feature heatmaps (d and e). Higher heterogeneity in radiomic expression observed through increasing proportion of blue/red within heatmaps of stricture wall annotation and within the 3 mm expansion band. (f) Receiver operating characteristic curves showing ML model performance for top ranked radiomic features (in both discovery and validation cohorts), rVAS score (in the entire cohort), and the augmented rVAS/radiomics model (in the validation cohort). (g) Bar plots for sensitivity, specificity, F1 scores, and accuracy of the models. (h) Average radiomic classifier probability for severe SMH plotted against ground truth histopathology pVAS. The central solid line shows the linear regression between these 2 variables with the thinner curves showing error bounds. CTE, computed tomography enterography; ML, machine learning; pVAS, pathologist’s Visual Analog Scale; rVAS, radiological VAS; SMH, smooth muscle hypertrophy/hyperplasia; TI, terminal ileum.

Combining radiologist assessment and radiomics yielded improved performance beyond radiological evaluation of pathological inflammation, fibrosis, or SMH on CTE

Based on standardized centrally read radiologist assessment (Table 2), strictures with pathologically severe inflammation had a median inflammation rVAS of 70 (range 10–100) whereas nonsevere inflammation strictures had a median rVAS of 70 (range 10–100), which were not found to be significantly different (P = 0.68). As shown in Figure 2g, this corresponded to a radiologist accuracy of 0.57 for inflammation rVAS in distinguishing histopathologically severe from nonsevere inflammation (sensitivity = 0.50, specificity = 1.00, F1 score = 0.68, AUROC = 0.43). Augmenting inflammatory rVAS with inflammation-associated radiomic features yielded a significantly improved performance compared with rVAS alone with a validation accuracy = 0.74 (sensitivity = 0.50, specificity = 1.00, F1 score = 0.67, AUROC = 0.66; P < 0.05), albeit not improved relative to radiomics alone.

Table 2.

Summary of rVAS of CTE scans within pathological groupings for severe/nonsevere fibrosis, SMH, and inflammation

		Pathological SMH			Pathological fibrosis			Pathological inflammation
Category/variable	All N = 100	Severe (N = 46)	Non-severe (N = 54)	P value	Severe (N = 47)	Non-severe (N = 53)	P value	Severe (N = 47)	Non-severe (N = 53)	P value
Inflammation rVAS Median (range)	70 (10–100)	60 (10–100)	70 (10–100)	0.1	60 (10–100)	70 (10–100)	0.01	70 (10–100)	70 (10–100)	0.68
Fibrostenosing rVASa Median (range)	30 (5–80)	30 (5–80)	30 (5–80)	0.16	30 (5–80)	30 (5–80)	0.73	30 (5–70)	30 (5–80)	0.70

CTE, computed tomography enterography; rVAS, radiological Visual Analog Scores; SMH, smooth muscle hypertrophy/hyperplasia.

^aSMH and fibrosis rVAS were evaluated against the same “noninflammatory changes” rVAS on CTE.

Strictures with pathologically severe fibrosis had median fibrostenosis rVAS of 30 (range 5–80) whereas nonsevere fibrosis strictures had a median rVAS of 30 (range 5–80), and were not found to be significantly different (P = 0.73, Table 2). Fibrostenosis rVAS also did not show significant differences when compared between severe and nonsevere SMH strictures, while yielding the same medians and ranges (P = 0.16). Figure 3g and Figure 4g show that the radiologist fibrosis rVAS resulted in an accuracy of 0.50 for distinguishing histopathologically severe from nonsevere fibrosis (sensitivity = 0.23, specificity = 0.81, F1 score = 0.32, AUROC = 0.51) and an accuracy of 0.44 in segregating histopathologically severe from nonsevere SMH (sensitivity = 0.17, specificity = 0.74, F1 score = 0.24, AUROC = 0.58). Augmenting fibrostenosis rVAS with fibrosis-associated radiomic features resulted in a markedly improved performance compared with rVAS alone for both fibrosis (validation accuracy = 0.68) and SMH subtypes (validation accuracy = 0.68).

DISCUSSION

Noninvasive characterization of inflammation, fibrosis, and SMH in patients with stricturing ileal CD remains critical to inform therapeutic choices (1), but cross-sectional imaging remains limited in its ability to accurately evaluate the severity of these subtypes (4). In this study, we showed that radiomics features capturing heterogeneity within and proximal to the bowel wall in terminal ileal CD strictures on CTE could accurately distinguish between each of severe and nonsevere histopathological inflammation, fibrosis, and SMH. Augmenting standardized radiologist scoring with radiomic features was also found to markedly improve performance relative to radiological visual assessment of strictures alone, for all 3 histopathological subtypes.

Recent studies have shown that muscular hyperplasia/hypertrophy, particularly in the submucosa and muscularis propria, is a dominant feature of stricturing segments which is correlated with chronic inflammation and is a key driver of luminal narrowing in fibrostenotic disease (2,10). Despite this, there are limited cross-sectional imaging parameters to help determine the degree of SMH or to distinguish fibrosis from SMH from inflammation, especially when considering widely available modalities such as CTE. A key finding of our work is the complementarity of radiomic features from regions proximal to annotated CD strictures when combined with bowel wall radiomic features, for characterizing pathological subtypes of fibrosis, inflammation, and SMH. In addition to distinct ripple, level, and edge patterns of bowel wall heterogeneity being associated with each histopathologically severe subtype, specific histogram and gradient heterogeneity descriptors from either a 3 mm (fibrosis, SMH) or 6 mm (inflammation) expansion were found to markedly augment machine classifier accuracy. This may be a result of radiomic features quantifying surrogates of predominant fibrosis characteristics occurring at more superficial tissue layers such as small bowel thickness, creeping fat, or fibrofatty proliferation vs deeper inflammatory or muscular changes in the muscularis propria and serosa (13), or perienteric edema (1,3). However, these findings require further validation to confirm the pathophysiological basis of radiomic features from the bowel wall and regions proximal to it, potentially through careful spatial mapping of imaging and pathology for a more detailed region-wise analysis (14).

Our results suggest the potential for radiomic features in characterizing pathological SMH and resonate with recent findings by Yu et al, albeit where the latter used a limited cohort of postcontrast MRI scans in conjunction with semiquantitative grading of histopathology (15). Our work also builds on previous radiomics studies examining both hand-crafted (6) and neural network models (16) for characterizing fibrosis through CTE scans, though the latter did not explicitly examine inflammatory and SMH components individually. Notably, our study involves a more rigorous approach to address potential shortcomings in previous work regarding cohort definition and selection of surgical specimens, where we used strict inclusion/exclusion criteria to derive a homogenous terminal ileal/neoterminal ileal CD cohort where CTE and histopathology surgical specimens could be spatially linked, used standardized CONSTRICT radiological definitions of CD strictures confirmed by a trained IBD radiologist (9), central reading by a highly experienced IBD radiologist trained by the STAR consortium, and consensus-based histopathological assessment scores from a trained IBD pathologist (3). Thus, we believe our cohort definition may better align with common sites for CD strictures while being more representative of patient populations who will be included in future clinical trials (17).

Our analysis of radiological visual assessment scoring found no significant differences between severe and nonsevere subtypes for inflammation, fibrosis, or SMH (with the same radiologist VAS score used for the latter two). This is in contrast to the literature (4,6,7,18,19), where radiological assessment has been found to be adequate fordetecting the presence of inflammation, less adequate for noninflammatory fibrostenosis, and not widely evaluated for SMH. Critically, our work suggests that while radiologist interpretation performs reasonably well for detecting the presence of inflammation, it may be substantially less reliable for grading severity and for distinguishing inflammatory from noninflammatory components of stricturing disease. In addition to reflecting the inherent constraints of qualitative imaging assessment, the difference in radiological performance relative to our own previous study (7) does suggest that the availability of additional MR parameters (contrast enhancement, diffusion) may enable more accurate radiological evaluation of CD strictures by magnetic resonance enterography (MRE) (7,20–23). Notably, we also had to account for additional technical challenges when developing ML-based radiomics models using CTE compared with our previous MRE-based models (7,24), potentially due to subtle differences as a result of acquisition-related or reconstruction-related differences across the cohort and poorer soft-tissue contrast and anatomic geometry on CTE scans. To address this, we implemented a more detailed reproducibility-informed framework to identifying radiomics features for each pathological subtype, which accounted for multiple sources of variation which resulted in improved overall accuracy for our machine classifier models (25). The use of volumetric, orientation-agnostic radiomic features and exclusion of geometry-dependent measurements further ensured robustness of our models across variable bowel configurations commonly encountered in stricturing CD. Interestingly, combining the radiologist VAS with radiomic features yielded markedly improved performance compared with radiologist VAS alone for all 3 subtypes in this study, albeit not significantly improved from using radiomics alone. In this study, radiologist VAS was used as a pragmatic and clinically relevant comparator, and we posit that more active and real-time integration of radiomics with radiological evaluation may reveal how best to integrate these complementary approaches for stricture characterization and interpretation on CTE scans.

We do acknowledge some limitations to our study. Despite stringent inclusion/exclusion criteria, changes in biologic medication occurred between CT imaging and surgical resection in 6 (out of 100) patients, whereas the remaining remained on a stable biologic treatment. Our overall cohort characteristics were relatively homogeneous otherwise. Inter-rater reliability was not evaluated in this study as only a single radiologist and a single pathologist were involved in the scoring. Future validation studies will include multiple assessors and potentially more active integration of radiomics to confirm clinical applicability. As histopathological scoring was conducted for only one representative cross-section from the narrowest area of the bowel, this may not account for heterogeneity in fibrotic, inflammatory, and SMH phenotypes along the entirety of a CD stricture. We are designing future prospective studies where multiple cross-sections can be matched between imaging and pathology, toward more detailed radiomic and radiologic characterization of pathological subtypes. We acknowledge that the definition of “severe” inflammation, fibrosis, and SMH in this study was based on cohort-specific pVAS distributions rather than validated clinical thresholds. Although this approach is common in exploratory radiomics and ML studies, it underscores the need for future multicenter efforts to establish standardized and clinically meaningful histopathological severity cutoffs. Our findings should therefore be interpreted as identifying relative, rather than absolute degrees of pathological severity within a surgically treated population. This study did not include longitudinal imaging-pathology pairs (acquired before and after advanced therapy) which limited our ability to study changes in severe inflammation, fibrosis, or SMH as a result of treatment. Future prospective cohorts with serial imaging and standardized therapy timelines will be critical to develop radiomics models for monitoring response and guiding treatment selection in stricturing disease. As our final cohort reflects an advanced degree of stricture severity requiring surgical resection, a significant proportion exhibited high severity of inflammation, SMH, and fibrosis (although this is a common feature of studies using resection as a gold standard). Nevertheless, our radiomics models yielded accurate signatures for distinguishing between less severe and more severe pathological subtypes, indicating their promise for quantifying biological signals associated with inflammation, fibrosis, and SMH within strictures on CTE. However, for these tools to meaningfully influence clinical decision making, further improvements regarding both sensitivity and specificity will likely be required. This will include additional optimization and evaluation of radiomics models for risk stratification or guiding the choice of interventions (medical escalation, surgery) in the context of larger, multicenter studies and prospective validation.

In summary, in this retrospective study of patients with stricturing CD, radiomic features on CTE provide a promising, noninvasive approach to characterize the dominant pathological components—inflammation, fibrosis, and SMH. Accurate differentiation of these components remains a major clinical challenge, yet is essential for individualized treatment decisions, including escalation of medical therapy, consideration of endoscopic dilation, or surgical intervention. We have started a larger multicenter investigation under the umbrella of the Stenosis Therapy and Anti-Fibrosis Research consortium that will also allow us to include a sufficient number of patients to distinguish degrees of fibrosis on a more granular level. Applying our tool for prognostication could allow the Stenosis Therapy and Anti-Fibrosis Research group to better assess outcomes in established stricturing disease toward triaging patients based on need for endoscopic dilation vs surgery. With further prospective validation and integration into real-time workflows, radiomics could bridge a critical gap in stricture evaluation, enhance prognostication, and improve patient stratification in clinical trials aimed at antifibrotic or antimuscular therapies.

Supplementary Material

SUPPLEMENTARY MATERIAL accompanies this paper at http://links.lww.com/AJG/D952

Study Highlights.

WHAT IS KNOWN

• Computed tomography enterography (CTE) is widely used for Crohn’s disease stricture evaluation.

• Differentiating fibrosis, inflammation, and smooth muscle hyperplasia within strictures is challenging.

• Radiomics can quantify imaging heterogeneity in Crohn’s disease.

• Previous CTE/magnetic resonance enterography radiomics focused mainly on fibrosis and inflammation.

WHAT IS NEW HERE

• Developed first CTE radiomics models for inflammation, fibrosis, and smooth muscle hyperplasia individually.

• Radiomics outperformed radiologist scoring for all three subtypes.

• Radiomic features from perienteric regions improved classification accuracy.

• Combining radiologist VAS and radiomics further enhanced performance.

• Demonstrated feasibility for noninvasive stricture subtyping in Crohn’s disease.

Data availability:

The data underlying this article will be shared on reasonable request to the corresponding author.