The records of patients with a diagnosis of CIS proven by conization and who underwent an MR scan between March 2013 and March 2016 in our hospital were retrospectively reviewed. A total of 110 patients were included to form the database in the present study. Clinical and pathological variables including age, parity, menopausal status, conization method, cone base area and depth, endocervical margin and glandular involvement, endocervical involvement based on ECC, and the number of quadrants with positive margins were collected. The study was approved by our Institutional Review Board.
The study was retrospectively designed and was carried out in two stages. First, from the database, we collected patients who ultimately underwent hysterectomy to form a cohort and establish an imaging model to predict the residual status after conization. Patients were randomly assigned to a training and testing group at a ratio of 1:1, and the performance of the imaging model was compared with the pathological positive margins in this stage. In the second stage, patients who opted for uterine preservation were included, and all patients were classified as high-risk or low-risk patients according to the imaging results. In this stage, the imaging model established in the first stage was used to classified the risk categories in patients with uterus preservation, and the performance of the model was evaluated. Patients with abnormal colposcopy or high-grade squamous intraepithelial lesion smear results in the follow-up procedure were subjected to a repeat biopsy, and the presence of histologically confirmed cervical intraepithelial neoplasia grades 2 or 3 (CIN2/3) or higher was considered residual or recurrent disease[17]. All patients were followed up for 24 months.
MRI protocol
All scans were performed using a 1.5 T MR scanner (Aera, SIEMENS, Erlangen, Germany) with the patient in the supine position. The following sequences were used to acquire images, from which features would be extracted for the radiomics model: axial T2-weighted imaging (T2WI) (repetition time/echo time=4500 ms/80 ms, slice thickness= 6 mm, gap = 1mm, field of view(FOV)= 320*240, flip angle= 160°, number of excitation= 2, with fat saturation), and axial diffusion-weighted imaging (DWI) (repetition time/echo time=5200 ms/80 ms, field of view = 250*200, slice thickness= 6 mm, gap = 1mm, flip angle= 90°, number of excitation= 6, b-value = 0, 800). The apparent diffusion coefficient (ADC) value was derived according to the following equation:
where S(b) and S(b0) represent the signal intensity of a certain voxel in the presence and absence of diffusion sensitization, respectively.
Area segmentation and radiomics feature extraction
This normalization approach has been used according to previous study [18], Three-dimensional volume of interest (VOI) of tumor contours were manually delineated slice-by-slice using the ITK-SNAP software, and VOI were first drawn to segment the uterine on the T2W image; then, based on the segmented uterine area, the cornization margin was delineated. Image erosion was applied to the binary segmentation mask for each cornization margin using a disk with a defined pixel radius, which was then eroded (disk radius of 3 pixels) to generate the VOI under cornization margin for further feature extraction. The VOI delineated on the T2W image was also applied to the ADC maps. The segmentation procedure is shown in Fig 1. For each segmented 3D volume, we extracted quantitative texture features from each phase using a program developed in-house. The texture features describe the high-order spatial distributions of intensities within the VOI. Fifty-two texture features were extracted from each sequence using several different methods, including the gray level co-occurrence matrix (GLCM), gray level run length matrix (GLRLM), gray level size zone matrix (GLSZM), and neighborhood gray-tone difference matrix (NGTDM). A detailed calculation of the texture features can be found in[15,18,19]. Finally, for each VOI, 156 features were extracted from the MR image. To find robust features against the intra- and interobserver delineation variations, the delineation was repeated on 40 patients by the same radiologist (G.W. with 10 years of experience in pelvic imaging) to assess intraobserver reliability and by another clinician (M.S. with 4 years of experience in pelvic imaging) to assess interobserver reliability. Parameters were included only when the agreement was good.
Statistical analysis
Intra- and interobserver agreement was analyzed based on the intraclass correlation coefficient (ICC), and a parameter with an ICC higher than 0.75 was considered to have good agreement[20]. All classification models were trained on the training cohort and tested on the independent test cohort. Both feature subsets selected with or without the Boruta method were analyzed. Multiple hypothesis correction was performed through a false discovery rate (FDR) adjustment using the Benjamini-Hochberg method[21]. The AUCs were statistically compared between different classifiers using the DeLong method[37]. All indices were calculated for both the training and test cohorts. For the validation cohort, the high-risk and low-risk patients classified according to the imaging model were compared and evaluated using disease-free survival (DFS) with the Kaplan-Meier curve. The statistical analyses were performed with R software(version 2.9.1).