Subject Selection
The ethics committee of our institutional approved this retrospective study(2020-KL-035-01) and the need for written informed consent was waived. Patients diagnosed with CD in two centers were retrospectively collected. All subjects were screened according to the inclusion and exclusion criteria. The inclusion criteria were as follows: (1) patients who performed ileocolonoscopy; (2) patients who underwent MRI scan. The exclusion criteria included: (1) Incomplete clinical data; (2) The interval between endoscopy and MRI was more than 3 days, in order to avoid the change of intestinal real-time state; (3) The area of interest (ROI) cannot be delineated or the TI is not clearly visible on imaging; (4) Patients whose TI was dominated by fibrous stenosis or surgically removed; (5) L2 sub-type and upper digestive tract involvement with L2 were excluded according to the Montreal Classification because of uninvolved TI.
Mri Acquisition
All patients in two centers were imaged using a 3.0T MR (General Electric Company, USA). Patients were placed in the supine position on the examination bed. An abdominal surface coil was used to modulate the signal. Before imaging, patients were given 20 mg Raceanisodamine Hydrochloride intramuscularly to reduce bowel peristalsis and dilate the lumen fully. The scanning sequences mainly included T2-weighted fat-suppression sequences, diffusion-weighted imaging sequences (b-values included 500 and 800mm2/s), and contrasted enhancement based on liver acquisition with volume acceleration (LAVA) sequence. Arterial-phase enhanced imaging based on liver acquisition with volume acceleration (LAVA) sequence was acquired 20 s after intravenous administration of 0.1 mmol/kg body weight of Gadolinium-DTPA through the cubital vein at a rate of 2.5mL/s. The detailed parameters are listed in Table1. Due to the application of the arterial phase in processing of radiomics, the scanning parameters of contrasted sequence in center 2 was only presented.
Table 1
Protocol for MR image acquisition
| Plane | TR/ms | TE/ms | Slice thickness/mm | FOV | Matrix | NEX |
T1WI+C(Center 1) | Axial/coronal | 3.7 | 1.1 | 5 | 280x80 | 288x288 | 1 |
T1WI+C(Center 2) | Axial | 3.7 | 1.6 | 5 | 280x80 | - | 1 |
T2WI-FS(Center 1) | Axial | 3333.3 | 85.2 | 7 | 36x80 | - | 2 |
DWI(500/800mm2/s,Center 1) | Axial | 7058.8 | 81.8 | 7 | 36x80 | - | 4 |
Note: TR: repetition time |
TE: echo time |
FOV: field of view |
NEX: number of excitations |
T1WI+C: Contrasted T1-weighted image |
T2WI-FS: T2-weighted image-Fat Suppression |
DWI: Diffusion weighted imaging |
Endoscopic Data And Maria Collection
Ileocolonoscopy was used as a reference standard of inflammatory bowel disease. All patients were given 3000-4000 mL of compound polyethylene glycol and electrolyte solution for intestinal cleansing the night before their examination, and to ensure the intestinal tract was clearly visible under endoscopy. 40ml Simethicone Emusion was also given orally on the morning of inspection. For comparing the consistency of the MaRIA and endoscopic assessment of the ROI, the TI, which was defined as the segment of the small intestine within 10 cm of the ileocecum, was assessed by CDEIS of terminal ileum(tCDEIS) rather than the overall CDEIS score. For CDEIS calculation, the endoscopic variables were as originally defined: deep ulcers and superficial ulcers (presence or absence), ulcerated surface and affected surface (evaluated on a 10 cm linear analogue scale), ulcerated and non-ulcerated stenosis [17]. All procedures were carried out by a gastroenterologist with more than 20 years of experience using the standard equipment (CFQ240, Olympus, Japan).
MaRIA was evaluated by MR findings including bowel thickness, relative contrast enhancement (RCE=((WSI postgadolinium − WSI pregadolinium) / (WSI pregadolinium)) × 100 × (s.d. noise pregadolinium / s.d. noise postgadolinium))), edema (hyperintensity on T2-weighted sequence relative to the signal of the psoas muscle) and ulcers (defined as deep depressions in the mucosal surface) [10]. Two radiologists, one with 15 years of experience in abdominal imaging (ER1) and one with 5 years of experience in abdominal imaging (ER2) were asked to calculate the score in the thickest region of the bowel wall according to these following formulas: MaRIA=1.5× bowel thickness + 0.02× RCE + 5× edema + 10× ulcer. The two radiologists were blind with each other and the endoscopic results.
MaRIA less than 7 indicates normal mucosa, a score of 7-11 is considered mild disease, a score greater than 11 indicates ulcerated lesions[11], and MaRIA less than 11 was treated as ulcer healing. In CDEIS score, 3.5 and below are classified as mild inflammation with a tendency toward mucosal healing, a score of 3.5-7 is considered moderated disease and less than 7 was treated as endoscopic remission, a score more than 7 was treated as ulcerative disease [18], which indicated a poorer prognosis and moderate-high risk of clinical treatment [19]. Therefore, all patients were classified as ulcerative group(UG, tCDEIS\(>\)7, MaRIA\(>\)11), and mucosal remission group (MG, tCDEIS\(\le\)7, MaRIA\(\le\)11).
Radiomics Processing
The arterial stage of T1 weighted enhanced imaging of all patients was imported into ITK-SNAP (Version 3.80) software in DICOM format. Two radiologists who evaluated MaRIA performed outlining along the TI wall on the axial image slice by slice manually. The lumen of intestinal was excluded and the intestinal wall was not clearly displayed or the visual overlap of adjacent intestinal walls were not included in the ROI. Part of the outlining results were shown in Figure 2A and Figure 2B.
All the outlining images and the corresponding primary images were imported into Dr. Wise Multimodel Research Platform (Version1.6.3.6 Deepwise&League of PHD Technology Co., Ltd, Beijing, China) for one-to-one matching, labeling and ROI features extraction. 1648 features were selected, including shape feature, Gray Level Co-occurrence Matrix (GLCM), Gray Level Run Length Matrix (GLRLM), Gray Level SizeZone Matrix (GLSZM), Gray Scale Correlation Gray Level Dependence Matrix (GLDM) and Neighborhood Gray-Tone Difference matrix (NGTDM), all these features were transformed according to LoG, Square, Square Root, Logarithm, Gradient, Wavelet and local binary pattern(LBP 2 dimension and 3dimension). All patients in center 1 were randomly divided into the training group and the experimental group in a ratio of 7:3 and underwent 5 cycles. Before feature dimension reduction, the features with a miss rate greater than 10% was selected to be cleared, and a correlation coefficient greater than 0.9 between features should be removed to avoid redundancy. Mutual information(MI) method was selected to control the multivariate and logistic model was selected for machine learning. The workflow in summary can be seen in Figure 2. The patients in center 2 was used to validate the model.
Calibration And Clinical Utility Of Radiomics Model
Calibration curve was used to measure how well a probabilistic prediction of an event matches the true underlying probability of the event. Decision curve analysis was used to measure the clinical efficacy of Radiomics model; A decision analysis measure, called the net benefit of the model, was calculated for the possible threshold probabilities. The benefits (proportion of true positives) and disadvantages (proportion of false positives) are added, and the diagnosis is weighted by the relative harm of false positives and false negative results. The net benefit values of the diagnostic model was standardized for prevalence.
Statistical Analysis
All statistical analyses were performed using SPSS software(version 22.0) and MedCalc software (version 15.2). Descriptive statistics were performed for part of the patient clinical data. Median(interquartile range) was performed to describe the distribution of data, such as tCDEIS, BWT, RCE, and MaRIA. The intraclass correlation coefficient(ICC) test and weighted Kappa coefficient were provided to test the consistency in measurements and features of ROI between radiologists. Area under the receiver operating curve (AUC), sensitivity and specificity were reported for presenting effectiveness of MaRIA evluated by radiologists, Delong test used to measure statistical differences between MaRIA and radiomics. A p-value less than 0.05 was considered statistically significant. Reliability was constitute “poor,” “fair,” “moderate,” and “good” reliability, with a corresponding cutoff value of ICCs or k value of <0.4, 0.41–0.6, 0.61–0.8, and >0.8, and higher than 0.9 was considered the good agreement in ROI.