Integrating Radiomics With Clinicoradiological Scoring Can Predict High-Risk Patients Who Need Surgery in Crohn’s Disease: A Pilot Study

Abstract

Background

Early identification of Crohn’s disease (CD) patients at risk for complications could enable targeted surgical referral, but routine magnetic resonance enterography (MRE) has not been definitively correlated with need for surgery. Our objective was to identify computer-extracted image (radiomic) features from MRE associated with risk of surgery in CD and combine them with clinical and radiological assessments to predict time to intervention.

Methods

This was a retrospective single-center pilot study of CD patients who had an MRE within 3 months prior to initiating medical therapy. Radiomic features were extracted from annotated terminal ileum regions on MRE and combined with clinical variables and radiological assessment (via Simplified Magnetic Resonance Index of Activity scoring for wall thickening, edema, fat stranding, ulcers) in a random forest classifier. The primary endpoint was high- and low-risk groups based on need for surgery within 1 year of MRE. The secondary endpoint was time to surgery after treatment.

Results

Eight radiomic features capturing localized texture heterogeneity within the terminal ileum were significantly associated with risk of surgery within 1 year of treatment (P < .05); yielding a discovery cohort area under the receiver-operating characteristic curve of 0.67 (n = 50) and validation cohort area under the receiver-operating characteristic curve of 0.74 (n = 23). Kaplan-Meier analysis of radiomic features together with clinical variables and Simplified Magnetic Resonance Index of Activity scores yielded the best hazard ratio of 4.13 (P = (7.6 × 10^-6) and concordance index of 0.71 in predicting time to surgery after MRE.

Conclusions

Radiomic features on MRE may be associated with risk of surgery in CD, and in combination with clinicoradiological scoring can yield an accurate prognostic model for time to surgery.

Key Messages.

• Patients with Crohn’s disease (CD) who are at risk for complications could benefit from targeted surgical referral, but existing clinical assessment and radiological evaluation have not been definitively associated with need for surgery.

• We identified a novel suite of computer-extracted image (ie, radiomic) features on routine magnetic resonance enterography that are associated with risk of surgery within 1 year of initiative aggressive medical therapy.

• Combining radiomics with clinical variables as well as radiological Simplified Magnetic Resonance Index of Activity evaluation could be used to prognosticate risk and need for surgery in CD, which could enable personalized treatment and targeted referral in CD patients with complications.

Introduction

Crohn’s disease (CD) affects over 1 million people in the United States alone, with a significant increase in prevalence over the past 5 decades.¹ Signs and symptoms of CD include abdominal pain, diarrhea, anemia, and malnutrition, while the disease presents as segmental inflammation of the gastrointestinal tract with a remitting and relapsing course.¹ While biologic² and combination³ therapies are commonly utilized to treat intestinal inflammation in CD, there continues to be a high rate of complications such as strictures and fistulas that require surgical therapy,⁴ most often in the first few years⁵ of starting treatment. A major challenge in CD treatment still remains identifying which patients are most at risk for disease complications and in whom early intervention can be highly effective for improving clinical remission rates.^6,7 Risk stratification is often determined based on clinical judgment and symptomatic disease activity,⁸ an imprecise and inaccurate methodology. This suggests a need for more accurate risk stratification of patients with CD,⁹ that is, early identification of those patients who should be referred to surgery to avoid CD complications from those who have a lower risk of complications and will respond to medical therapy (biologics, immunomodulators, or a combination) without requiring surgery.¹⁰

Routine clinical evaluation of CD typically includes laboratory assessment for inflammatory markers,¹¹ Montreal classification,¹² endoscopy, and cross-sectional magnetic resonance enterography (MRE) or computed tomography enterography imaging.¹³ Activity indices such as the Crohn’s Disease Activity Index or the Harvey-Bradshaw index (HBI) have also been utilized to aggregate clinical assessments¹⁴ into singular scores, albeit with poor predictive ability¹⁵ and only moderate performance in predicting need for CD-related surgery.⁸ Cross-sectional imaging via MRE has demonstrated excellent sensitivity and specificity for active CD diagnosis¹⁶ and moderate performance in identifying CD complications (strictures, ulcerations)¹⁷ that are associated with risk of surgery.¹⁸ The need for more objective imaging assessment of CD has resulted in development of approaches¹⁹ such as the Magnetic Resonance Index of Activity²⁰ or simplified Magnetic Resonance Index of Activity (sMARIA),²¹ the Clermont score, the Magnetic Resonance Enterography Global Score, and the Lémann index, which leverage expert measurements of wall thickness and wall enhancement in addition to an assessment of other CD-related complications. These scores have shown some association¹⁹ with active CD or CD response to therapy but have not been definitively correlated with the need for surgery²² or predicting risk of complications. Given the significant anatomic detail that can be appreciated on MRE scans,²³ there may thus be an opportunity to investigate and develop more advanced quantitative imaging measurements that can be associated with CD.

Advances in the field of radiomics (ie, quantitative imaging descriptors from radiographic imaging)²⁴ have shown great promise for deriving computational imaging measurements that are associated with treatment response and patient prognosis in oncology. Radiomic features have been shown to quantify subtle details on routine magnetic resonance imaging (MRI) scans, which may not be visually appreciable, and which can be correlated with underlying pathophysiology²⁴ as well as with disease response to treatment.²² Initial studies of radiomic features in CD have demonstrated that they may be able to capture histological disease activity^25,26 as well as characterize bowel fibrosis,²⁷ suggesting their potential to prognosticate need for surgery in CD.

In this study, we constructed and evaluated a prognostic radiomics model to risk-stratify patients with CD based on need for early surgery after aggressive medical therapy (biologics, immunomodulators, or a combination) using retrospectively accrued routine MRE scans. We hypothesized that the combination of radiomic features with clinical variables (disease behavior, serum markers, and demographics) as well as radiological assessment (based on sMARIA scoring for presence of wall thickening, edema, fat stranding, and ulcers) could more accurately distinguish high-risk from low-risk patients and predict time to surgical intervention after medical therapy for CD.

Methods

Study Population

This was a single-center, retrospective study of patients at University Hospitals Cleveland Medical Center in Cleveland, Ohio. After Institutional Review Board approval, electronic medical records were reviewed to identify patients treated for CD between 2009 and 2015 (as coded by International Classification of Diseases–Ninth Revision codes, category 555) after referral to either of 2 primary adult gastroenterologists (J.A.K. or M.D.), with patient outcomes determined via review of patient charts up to 2020. Inclusion criteria were that these patients be over 18 years of age; have been recommended a biologic (adalimumab, infliximab, or certolizumab pegol), immunomodulator (6-mercaptopurine, azathioprine, methotrexate), or combination (biologic + immunomodulator therapy) at the point of study inclusion; and have undergone an MRE exam within 3 months of study inclusion and prior to starting treatment. Patients who did not have MRE scans or who had isolated jejunal/duodenal disease without ileal disease were excluded. Two physicians (M.C., A.S.) independently abstracted the data, and discrepancies were resolved by consensus with the senior clinical investigator (M.D.). For all included patients, clinical variables recorded were sex, body mass index, smoking history, erythrocyte sedimentation rate, C-reactive protein, type and duration of medication, and surgical history. Additionally, disease location, upper tract involvement, disease behavior, and perianal involvement were used to determine Montreal classification.²⁸ The primary patient outcome was defined based on need for and timing of surgery to construct 2 risk groups: high risk (patient underwent surgery within 12 months of MRE imaging) and low risk (patient did not have surgery or had surgery 13+ months after MRE imaging).

MRI and Annotation

From the MRE protocol used at University Hospitals Cleveland Medical Center, the sequence that was commonly available across all patients was a T2-weighted half-Fourier acquisition single-shot turbo spin echo (repetition times = 500-971 ms; echo times = 37-80 ms) scan acquired in the axial plane after contrast agent administration on a 1.5T or 3T scanner (Siemens Avanto, Siemens Verio scanner, Phillips Ingenia). Table 1 summarizes imaging parameters for all MRI scans used in this study. Based on available radiological reports, a radiologist (K.B.) with expertise in abdominal cross-sectional MRI first identified the ileocecal junction (between vertebrae L2-L4 on the MRI scan) and annotated the bowel wall for 3 to 5 cm above this landmark. Additionally, they scored the terminal ileum (TI) segment on each MRE scan based on sMARIA guidelines (ie, presence of edema, fat stranding, wall thickening, and ulcers).²¹ Total annotation time was estimated to be between 5 and 10 minutes per patient, while radiological sMARIA scoring required an additional 10 to 15 minutes per scan, all of which was conducted while blinded to risk groupings.

Table 1.

Summary of magnetic resonance enterography imaging parameters and scans used in this study

Imaging Parameter	3T (n = 20)	1.5T (n = 53)
Image quality
In-plane resolution, mm	0.78-1.5	1.17-1.95
Slice thickness, mm	4 or 5	4 or 5
Repetition time, ms	600-1100	500-900
Echo time, ms	45-80	39-52
Scanner
Siemens Avanto	0	49
Siemens Espree	0	1
Siemens Verio	6	0
Philips Achieva	1	0
Philips Ingenia	13	0
Unknown	0	3

Postprocessing and Radiomic Feature Extraction:

The overall experimental workflow and radiomic analysis is summarized in Figure 1. A publicly available MRI quality control package was used to determine which artifacts were present in the imaging cohort.²⁹ While no cases were excluded, image quality measurements demonstrated presence of bias field inhomogeneity (variation in intensity appearance across the volume)³⁰ and significant variations in voxel resolution (in-plane = 0.76-1.95 mm; slice thickness = 4-5 mm). These were accounted for via: (1) trilinear interpolation to resample the images to 1.3-mm in-plane and 5-mm slice thickness (determined as the mode resolution across the cohort) and (2) N4ITK Bias Field Correction³¹ in 3D Slicer to correct bias field inhomogeneity artifacts.

Figure 1.

Overall experimental workflow for integrating radiomics, clinical variables, and radiological assessment to develop a prognostic model for time to surgery in Crohn’s disease via magnetic resonance (MR) enterography scans. AUC, area under the curve; HR, hazard ratio; sMARIA, Simplified Magnetic Resonance Index of Activity.

A total of 183 2-dimensional radiomic features were extracted on a per-pixel basis across all TI slices of postprocessed MRI volumes to capture subtle variations in image appearance and heterogeneity^24,32 within the bowel wall based on responses to different texture operators. Supplementary Table S1 provides International Biomarker Standardization Initiative–compliant³³ definitions for all radiomic descriptors utilized in this study. Statistical descriptors (mean, variance, kurtosis, and skewness) were then calculated across the entirety of the TI subvolume, yielding a total of 732 radiomic descriptors associated with each patient.

Statistical Analysis

Categorical clinical variables and component sMARIA assessments were described as proportions or percentages and statistically compared via a chi-square test. Continuous serum markers as well as the composite sMARIA score were described as mean or median (including suitable measures of variability) and compared via nonparametric Wilcoxon rank sum testing. The criterion for statistical significance was P < .05.

Radiomic descriptors associated with need for surgery within 1 year of MRE were identified via the machine learning approach described in Supplementary Section S1. The performance of selected radiomic descriptors in discriminating between risk groupings was evaluated within both a discovery set (comprising a randomly selected set of 67% of patients considered) and hold-out validation set (comprising the remaining patients) with the proportions of low- and high-risk patients maintained across all cohorts. Performance of top-ranked radiomic descriptors and clinical variables and sMARIA components were evaluated via the area under the receiver-operating characteristic curve (AUROC) of the associated random forest classifier models, both individually and in combination. Wilcoxon rank sum testing was used to assess significant differences in area under the curve performance between feature sets.

Kaplan-Meier analysis was conducted to evaluate the performance of different feature sets (radiomic descriptors, clinical variables, sMARIA, combined feature sets) in predicting time to surgery during the entire follow-up period in a cross-validation setting. Kaplan-Meier curves were statistically compared between feature sets using a log-rank test as well as in terms of hazard ratios (HRs) and concordance indices (C-indices), described in Supplementary Section S2. Additional subgroup analysis of different Montreal classifications was conducted to determine the value-add of different feature sets for decision making in the presence of disease complications, and quantified via confusion matrices. Finally, radiomic descriptors were evaluated for robustness to differences in magnetic field strength, scanner manufacturer, and region of interest. Radiomic descriptors were also correlated against individual components of the sMARIA score (via Pearson correlation) to gain an understanding of their radiological interpretability. All radiomic and statistical analysis was conducted using MATLAB 2021b (The MathWorks, Inc., Natick, MA, USA). All authors had access to the study data and reviewed and approved the final manuscript.

Results

Study Population

A total of 80 patients met all inclusion criteria. Three patients were excluded due to lack of available T2-weighted MRI scans, while another 4 patients could not have a Montreal classification assigned due to missing information in patient charts. The final cohort comprised 73 adult patients with CD, of whom 21 patients were determined to be high risk (received surgery within 12 months of MRI), while 52 patients were found to be low risk (no surgery or surgery 13+ months after MRE). Demographic and clinical characteristics of both risk groups are summarized in Table 2.

Table 2.

Summary of demographics, clinical variables (serum markers, disease behavior), and treatment types for of patients included in this study

Category/Variable	Groups/Statistic	All (N = 73)	High Risk (n = 21)	Low Risk (n = 52)	P Value
Sex	Male	35 (47.9)	10 (47.6)	25 (48.1)	.97
	Female	38 (52.1)	11 (52.4)	27 (51.9)
Age, y	Mean ± SD	36.3 ± 13.7	34.1 ± 13.3	37.1 ± 13.8
	Median (range)	32.70 (16.7-71.1)	31.65 (16.7-71.1)	32.94 (19.7-66.9)	.38
BMI, kg/m2	Mean ± SD	28.1 ± 7.1	27.5 ± 6.8	28.3 ± 7.3
	Median (range)	26.60 (16.3-49.6)	25.43 (16.3-49.6)	26.95 (19.8-45.7)	.34
CRPa, mg/l	Mean ± SD	2.2 ± 2.1	1.1 ± 2.6	2.1 ± 1.4
	Median (range)	1.39 (0.3-9.4)	1.07 (0.3-9.4)	1.52 (0.3-5.15)	.51
	Not available	31 (42.5)	22 (30.1)	9 (12.3)
	CRP <0.3	22 (30.1)	5 (23.8)	17 (32.7)
ESR, mm/hr	Mean ± SD	22.2 ± 21.8	23.1 ± 23.7	21.4 ± 20.6
	Median (range)	14.50 (1-86)	13.00 (2-86)	16.00 (1-62)	.85
	Not available	37 (50.1)	22 (30.1)	12 (16.4)
Immunomodulators	6-MP	12 (16.4)	3 (14.3)	9 (17.3)
	Azathioprine	11 (15.1)	3 (14.3)	8 (15.4)	.60
	Methotrexate	3 (4.1)	0 (0.0)	3 (5.8)
Biologics	Humira	19 (26.0)	6 (28.6)	13 (25.0)
	Cimzia	2 (2.7)	1 (4.8)	1 (1.9)	.83
	Remicade	14 (19.2)	4 (19)	11 (21.2)
Combination	Immunomodulators + biologics	11 (15.1)	7 (9.6)	4 (5.5)	NA
Surgery	Yes	33 (45.2)	NA	12 (23.1)	NA
	No	40 (54.8)	NA	40 (76.9)
Location	L1, terminal ileum	28 (38.4)	6 (28.6)	22 (42.3)
	L2, isolated colon	3 (4.1)	1 (4.8)	2 (3.8)	.69
	L3, ileocolon	40 (54.8)	13 (61.9)	27 (51.9)
Upper tract Involvement	L4	3 (4.1)	1 (4.8)	2 (3.8)	NA
Behavior	B1, nonstricturing nonpenetrating	17 (23.3)	2 (9.5)	15 (28.8)	.03
	B2, stricturing	46 (63.0)	13 (61.9)	23 (44.2)
	B3, penetrating	10 (13.7)	6 (28.6)	4 (7.7)
Perianal	p, perianal	9 (12.3)	3 (14.3)	6 (11.5)	NA
Smoking	Current and past	30 (41.1)	9 (42.9)	21 (0.40)	.84
	Never	43 (58.9)	12 (57.1)	31 (59.6)

Values are n (%), unless otherwise indicated.

Abbreviations: 6-MP, 6-mercaptopurine; BMI, body mass index; CRP, C-reactive protein; ESR, erythrocyte sedimentation rate; NA, not applicable.

^aComputed without including values <0.3, as exact values for those patients were unavailable.

Clinical Characteristics and sMARIA Assessment

Disease location was predominantly ileocolonic (54.8%), while complicated stricturing or penetrating disease was present in 76.7% of the patients. A total of 49.3% of patients received biologics after MRI, of which 11 (15.1%) were on combination therapy. For risk groups defined based on need for surgery, 40 (76.9%) low-risk patients did not undergo any form of surgery in long-term follow-up (up to 60 months after initial study inclusion). Among patients who underwent surgery, low-risk patients had a median time to surgery of 24 ± 11.3 months vs 6 ± 2.9 months for high-risk patients (P = 2.3 × 10^-13). No significant differences were observed between high- and low-risk patient groups in terms of age, sex, body mass index, serum markers (C-reactive protein, erythrocyte sedimentation rate), history of prior surgery, or disease location. An additional breakdown of the types of surgery undergone by patients is provided in Supplementary Table S2. Montreal behavior (P = .03) was found to be significantly different between risk groups in which a larger proportion of high-risk patients presented with stricturing disease (61.9% vs 44.2%) or penetrating disease (28.6% vs 7.7%). Smoking history was also found to be significantly different between risk groups (P = .04), with a greater proportion of high-risk patients (42.9%) being current or past smokers, while more low-risk patients (59.6%) had never smoked.

As summarized in Table 3, the composite sMARIA score was found to be significantly higher in high-risk compared with low-risk patients (P = .002). A significantly higher proportion of high-risk patients were assessed with mural edema (61.9% vs 21.2%), increased wall thickness (81% vs 50%), and fat stranding (33.3% vs 8%). No significant differences were observed in radiological assessment of ulceration, but this was present in only 12.3% of patients overall.

Table 3.

Summary of sMARIA radiological assessment for magnetic resonance enterography scans included in this study

Category/Variable	Groups/Statistic	All (N = 73)	High-Risk (n = 21)	Low-Risk (n = 52)	P Value
sMARIA	Mean ± SD	1.31 ± 1.45	2.14 ± 1.56	0.98 ± 1.28
	Median (range)	1 (0-5)	2 (0-5)	0.5 (0-4)	.002
Wall thickness >3 mm	Yes	43 (58.9)	17 (81.0)	26 (50.0)
	No	30 (41.1)	4 (19.0)	26 (50.0)	.015
Mural edema	Yes	24 (32.9)	13 (61.9)	11 (21.2)
	No	49 (67.1)	8 (38.1)	41 (78.8)	7.9 × 10-4
Fat stranding	Yes	11 (15.1)	7 (33.3)	14 (26.9)
	No	62 (84.9)	14 (66.7)	38 (73.1)	.0056
Ulcers	Yes	9 (12.3)	4 (19.0)	5 (9.6)
	No	64 (87.7)	17 (81.0)	47 (90.4)	.28

Values are n (%), unless otherwise indicated.

Abbreviation: sMARIA, Simplified Magnetic Resonance Index of Activity.

Radiomic Descriptors Associated With Risk of Surgery

Eight radiomic descriptors that were most frequently identified as being significantly different between high- and low-risk groups are summarized in Table 4. A total of 62.5% of these features (n = 5 of 8) are gradient organization responses capturing entropy (chaos or disorder) or correlation (local texture) within the TI on MRE scans. Figure 2 visualizes heatmaps of 2 representative radiomic descriptors for high-risk and low-risk patients within a 2-dimensional TI section. Figures 2B and 2C show that low-risk patients exhibit less negatively skewed texture responses (seen as mixture of blue and red regions) compared to high-risk patients. By comparison, normalized T2-weighted signal intensity values (Figure 2A) do not show significant differences between risk groups. Top-ranked radiomic descriptors were also found to be highly robust to differences in magnetic field strengths as well as to scanner manufacturer, as described in Supplementary Section S3. As summarized in Supplementary Section S4, several radiomic descriptors were found to exhibit a weak correlation with specific sMARIA components of wall thickening, fat stranding, and ulceration, as well as being most strongly correlated with sMARIA-scored edema.

Table 4.

Most frequently selected radiomic descriptors associated with risk of surgery in Crohn’s disease

Radiomic Descriptor	High Risk (n = 21)	Low Risk (n = 52)	P Value
CoLlAGe sum entropy ws = 5	1.030 ± 1.962	-0.669 ± 1.409	.001
Gray median ws = 11	0.738 ± 1.220	-0.117 ± 1.649	.008
Haralick inverse difference Moment ws = 11	-1.138 ± 2.090	0.185 ± 1.696	.004
Laws ripple-wave ws = 5	0.717 ± 1.589	-0.050 ± 1.483	.021
CoLlAGe correlation ws = 5	0.666 ± 1.398	-0.157 ± 1.579	.016
CoLlAGe sum average ws = 3	-0.957 ± 1.058	0.526 ± 1.533	.005
CoLlAGe sum average ws = 5	-0.855 ± 1.864	0.039 ± 1.450	.005
CoLIAGe energy ws = 3	0.178 ± 1.334	-0.400 ± 1.583	.063

Values are mean ± SD.

Figure 2.

Representative radiomic heatmaps overlaid within annotations of the terminal ileum on T2-weighted magnetic resonance imaging scans, depicting subtle differences between Crohn’s disease patients who did not get surgery or had surgery 13+ months after magnetic resonance enterography (left, low risk) and those who got surgery within 1 year of magnetic resonance enterography (right, high risk). Normalized intensities (top row) are not seen to be significantly different between risk groups (P > .05). By contrast, top-ranked radiomic features exhibit significant differences (P = .001 and P = .004) in local image heterogeneity between low-risk and high-risk patients.

Risk Prediction for Time to Surgery

Radiomic descriptors, both independently as well as in conjunction with sMARIA scores and clinical variables, were able to accurately discriminate between risk groups in the discovery set (radiomics: 0.67 ± 0.08; combined: 0.69 ± 0.09; via an random forest model). Notably, the combination of radiomics with sMARIA and clinical variables yielded the best overall performance in hold-out validation (AUROC, 0.83 vs 0.74 for radiomics alone) (Figure 3A). By comparison, clinical variables (discovery AUROC, 0.48 ± 0.07) as well as sMARIA (discovery AUROC, 0.58 ± 0.07) yielded significantly lower performance for discriminating between high- and low-risk patients, a trend also observed in the validation set (AUROC clinical, 0.63; sMARIA AUROC, 0.77). Relative differences in classifier performance between radiomic descriptors, clinical variables, sMARIA scoring, and their combination is illustrated via the differently colored receiver-operating characteristic curves in Figure 3A. Kaplan-Meier analysis demonstrated a corresponding significant improvement in predicting time to surgery when considering radiomics descriptors or sMARIA alone (HRs, 3.49 and 3.68, respectively) (Figure 3C, D) compared with clinical variables alone (HR, 1.66) (Figure 3B). The combination of clinical variables and radiomic descriptors was found to yield the best overall performance in predicting time to surgery with an HR of 4.13 and C-index of 0.71 (Figure 3E).

Figure 3.

A, Receiver-operating characteristic curves comparing model performance between clinical variables (green), radiomics (blue), Simplified Magnetic Resonance Index of Activity (sMARIA) (red), and combined (black). Corresponding Kaplan-Meier curves for different feature sets based on median risk predictions from random forest models based on (B) clinical variables, (C) sMARIA scores, (D) radiomic descriptors, and (E) combined radiomics + sMARIA + clinical features. AUROC, area under the receiver-operating characteristic curve; CI, concordance index (See online version for color figure).

Figure 4 summarizes subgroup analysis for different predictors in distinguishing between high- and low-risk patients for Montreal classifications (B1, B2, and B3), via confusion matrices and receiver-operating characteristic curves. Combining clinical variables, sMARIA scores, and radiomic descriptors yielded the best overall accuracy in classifying risk groups within B1, B2, and B3 classifications while also identifying 100% of high-risk patients in each case. By contrast, the clinical model performed the worst in distinguishing risk groups within each Montreal classification, while sMARIA scores and radiomic descriptors performed comparably in all subgroups.

Figure 4.

Subgroup analysis of predictor performance within different Montreal subclassifications: (top) no complications (B1), (middle) stricturing (B2), and (bottom) penetrating (B3). Receiver-operating characteristic curves and confusion matrices depict differences in model performance between clinical variables, Simplified Magnetic Resonance Index of Activity (sMARIA) scores, radiomic descriptors, and combined radiomics + sMARIA + clinical features. FPR, false positive rate; ROC, receiver-operating characteristic; HR, hazard ratio; TPR, true positive rate.

Discussion

In this retrospective pilot study of patients with CD, radiomic features extracted from TI regions on MRE scans were found to accurately segregate patients into high-risk and low-risk groups based on need for surgery within 1 year of imaging. Combining radiomic features with clinical variables (disease behavior, serum markers, and demographics) as well as with radiological assessment (based on sMARIA scoring for presence of wall thickening, edema, fat stranding, and ulcers) resulted in a highly accurate multivariate prognostic model for time to surgery in CD.

Currently, there are a lack of well-validated noninvasive markers to stratify patients with CD into risk categories,⁹ especially to identify patients likely to suffer severe disease complications that are best treated via surgical intervention.⁷ In this study, radiomic features that captured texture responses from within the bowel wall were found to overexpress in high-risk patients who required surgery earlier in their disease course. These findings parallel previous studies in oncology in which this class of feature responses has been overexpressed in tumors less likely to respond to therapy,³² as well as in patients with poorer survival despite treatment.³⁴ While one of the top-ranked radiomic features was the median intensity within the bowel wall (known to be associated with acute inflammation and edema and also found to be associated with radiological assessment of edema in our experiments),³⁵ a majority of identified radiomic features captured subtle intensity and gradient differences in local neighborhoods within the diseased bowel wall. These texture responses may be surrogate measurements of underlying inflammatory or CD disease activity,³⁶ as suggested by previous studies in CD.^25,26 Correlative analysis in our study further validated this observation, as several radiomic descriptors were found to be weakly associated with radiological hallmarks of CD including wall thickening, edema, fat stranding, and ulceration.

The primary clinical variable found to be associated with risk of surgery was disease behavior (assessed via Montreal classification), in which stricturing and penetrating disease was associated with higher-risk patients, as borne out by several large-scale studies in CD.^37,38 While other clinical assessments (eg, HBI, stool frequency)⁸ have been found to be associated with risk of surgery in CD, these variables were unavailable in the current study. Conversely, while smoking has typically been considered a negative prognostic factor in CD,³⁹ omission of smoking history from our clinical model yielded no significant change in prognostic performance. As our study primarily included patients treated between 2009 and 2015, the results reflect the natural history of CD with biologics only utilized in ~50% of patients. Notably, the combination of clinical variables, radiological sMARIA scoring, and radiomic descriptors yielded the best overall accuracy for distinguishing between risk groups in all Montreal classifications in subgroup analysis, including predicting 100% of high-risk CD patients. This may be due to the combined predictor better accounting for false negative errors among high-risk patients compared with the individual feature sets. This attests to the significant potential of radiomics in CD, especially in conjunction with MRE, which is increasing in popularity¹⁶ due to the lack of ionizing radiation and the increasing need to do multiple follow-up scans. Our results also resonate with recent studies¹⁸ in which imaging evaluation of stricturing, wall thickness, and penetrating disease were found to be associated with risk of surgery in CD, though our multivariate model did yield improved AUROC performance across segregated discovery and hold-out validation cohorts. Our pilot study thus suggests that combining radiomics with clinical features and radiological scoring may enable improved classification and prognostic performance in predicting risk of surgery in CD.

The work presented in this study did have its limitations. The reported results were based on patients accrued from a single institution with segregation into discovery and hold-out validation cohorts. The number of patients available for analysis was further limited due to the stringent inclusion and exclusion criteria to ensure that the natural history of CD treatment did not impact our analysis. Despite detailed chart review, clinical activity indices such as HBI or Crohn’s Disease Activity Index as well as endoscopic scoring were not available for a majority of patients in this study and thus could not be included. Other symptoms that have been found to be associated with need for surgery such as abdominal pain and obstructive symptoms¹⁸ were not consistently available in our retrospective chart review and will be incorporated in a future prospective study. However, our study did evaluate sMARIA scoring of MRE scans for predicting risk of surgery in CD, for which follow-up was collected for up to 5 years. This allowed us to conduct one of the first integrated evaluations of clinical variables, radiological scoring, and radiomic descriptors for predicting time to surgery in CD. We plan to conduct a future prospective study in which our integrated predictor can be validated for prognosticating time and need for surgery in patients newly diagnosed with CD. Our pilot study only utilized a single reader for blinded annotations as well as radiological scoring of axial MRI scans, which allowed us to avoid issues related to the moderate agreement between observers in MRE interpretation of CD.⁴⁰ However, in addition to systematically accounting for multiple sources of image artifacts in MRE scans (including bias field and resolution differences) prior to radiomic analysis, we conducted a detailed evaluation of robustness of the top-ranked radiomic descriptors to sources of variation (magnetic field strength, scanner) in order to confirm their generalizability. Evaluating the interplay of radiomic descriptors and interobserver variability in radiological scoring will be an avenue for future work in a multicenter prospective study.

Conclusions

In summary, in this retrospective study of CD patients who underwent surgery, radiomic features from the diseased bowel wall on MRE scans were found to be significantly associated with risk of surgical intervention. The complementarity of radiomic features to clinical variables and radiological scoring was demonstrated by the combined model achieving the best overall prognostic performance for predicting time to surgery in patients with CD. Noninvasive risk stratification of patients with CD via a multivariate predictor integrating quantitative radiomic markers, clinical variables, and radiological scoring on routine MRE scans could enable more targeted treatment paradigms and hence better overall patient outcomes in CD.

Author Contribution

All authors approved the final version of the article, including the authorship list. Guarantors of this article are M.D. and S.E.V. P.C. performed data analysis, figure visualizations, and initial draft writing. A.S. and H.M.C. performed detailed chart review (including follow-up) and extracted clinical variables. K.B. assisted in obtaining, annotating, and scoring magnetic resonance imaging (MRI) scans under the supervision of R.P. J.A.K. and M.-J.R. assisted in statistical analysis and writing/editing of the manuscript. K.A. assisted in quality evaluation and radiomic analysis of MRI scans. A.M. and C.L. assisted in project conception, experimental design, and editing the final manuscript. R.P. provided access to MRI scans, supervised annotation/scoring of MRI scans, and provided input on imaging analysis. S.L.S. and J.A.K. provided clinical input on experimental design, and assistance in editing the manuscript. S.E.V. and M.D. jointly conceived and supervised the project, designed the experiments, oversaw the analysis, and wrote and edited the final manuscript.

Funding

This work was supported by the National Institute of Diabetes and Digestive and Kidney Diseases (1K08DK110421-01, 5P30DK097948-07, 1F31DK130587-01A1), the National Cancer Institute (1U01CA248226-01), the Peer Reviewed Cancer Research Program (W81XWH-21-1-0345), and the University Hospitals’ Research and Education Institutes Pilot Award. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health, the Department of Defense, or the U.S. Government.

Conflicts of Interest

A.M. is an equity holder in Picture Health, Elucid Bioimaging, and Inspirata; has served on the advisory board of Picture Health, Aiforia, and SimBioSys; has served as a consultant for Biohme, SimBioSys, and Castle Biosciences; holds research agreements with AstraZeneca, Boehringer Ingelheim, Eli-Lilly, and Bristol Myers Squibb; has had technology licensed to Picture Health and Elucid Bioimaging; and has been involved in 3 different R01 grants with Inspirata. S.E.V. has received research funding from Pfizer. The other authors disclose no conflicts.

Data Availability

Code used for radiomic analysis will be made available via GitHub. Data may be made available upon reasonable request and pursuant to appropriate agreements.