A radiomic nomogram based on arterial phase of CT for differential diagnosis of ovarian cancer

Our institutional review board approved this retrospective study with a waiver of informed consent. We reviewed electric database of POC and SOC from January 2010 to December 2018 and retrieved the images of enrolled patients from the picture archiving and communication system (PACS) system. The exclusion criteria of this study are as follows: (1) Patients with poor image quality; (2) patients without enhanced scanning; (3) patients with unclear boundary and unable to outline. Eligible patients including 62 POC patients and 48 SOC patients were randomly divided into the training cohort and the validation cohort. We collected 48 SOC patients, including 20 metastasis of gastric cancer, 15 metastasis of colorectal cancer, and 13 metastasis of sigmoid colon cancer. Among the 48 SOC patients, 18 patients were found to have ovarian masses before the primary malignancy revealed. The post-processing process of the image is shown in Fig. 1, and the flow chart of the whole study is shown in Fig. 2.

Preoperative abdominal CT scans were obtained using a Brilliance ICT (Philips Medical Systems, the Netherlands). The scanning protocol includes unenhanced and contrast-enhanced CT with arterial phase after intravenous administration of iodine-contrast agent (320 mg I/mL) at a rate of 3.0 mL/s with a dose of 2 mL/kg using an automated power injector. The scan parameters were tube voltage of 120 kVp, a pitch value of 0.99, a matrix of 512 × 512, slice thickness and interval both of 5 mm, and milliamperage was adjusted automatically according to the patient's size (ranged between 220 and 400 mA).

The contrast-enhanced CT images of enrolled patients were exported in Digital Imaging and Communication in Medicine (DICOM) format. Then ITK-SNAP was employed for tumor segmentation. We used the polygon mode to delineate cystic and solid component of tumors at each slice manually. All segmentations were conducted by a radiologist (C.L.) who had 16 years of imaging diagnosis experience and re-checked by H.Z. (15 years of experience).

Radiomics features including Histogram, Formfactor, Grey-level size zone matrix(GLSZM), and Run-length matrix (RLM) were extracted by AK software (Artificial Intelligence Kit V3.0.0. R, GE Healthcare, China) [13]. The preprocessing before feature selection was divided into three steps. Firstly, we sought to identify the features that contribute to the model using the ANOVA + KW test. And then we used the binary logistic regression analysis to rule out features, in which the correlation coefficient was greater than 0.9. Finally, the LASSO Cox logistic regression model was used to select the most useful prediction features. Then, the radiomics score (Rad-score) was computed for each patient through a linear combination of selected features weighted by their respective coefficients [14].

Multivariate logistic regression analysis was used to evaluate the significant clinical factors for distinguishing POC and SOC. Radiomics signature was applied to develop a distinguish model by using the training cohort. Then we combine the rad-score with clinical indicators and established a combination model. The normal reference range for CEA values is 0–10U/ML, for CA125 value is 0–35U/ML, and for CA199 value is 0–37U/ML. Finally, we constructed a visualized nomogram based on the combination model. The calibration curve was used to assess the nomogram. The Hosmer–Lemeshow test was performed to evaluate the goodness-of-fit of the nomogram.

Continuous and categorical variables were compared using the t test and chi-square test, respectively. Multivariable logistic regression analysis was used to select the independent prognostic factors. The performance of the model was assessed in the primary and validation cohorts. The discrimination of the signature was measured by the area under the curve (AUC). Statistical analysis was performed with SPSS (version 19.0, IBM, Armonk, NY, USA). A two-sided p value was always computed, and a difference was considered significant at P < 0.05.

A total of 110 ovarian cancer patients were eligible for this study. The 110 patients were divided into a training cohort (N = 76) and a testing cohort (N = 34) (Fig. 2). The baseline of enrolled patients in training and validation cohorts is displayed in Table 1. There were 56.6% (43/76) of POC patients in the training cohort and 55.9% (19/34) in the validation cohort. The mean age of the patients in the training cohort was 54.74 ± 11.01 years for patients with POC and 50.79 ± 10.24 years with SOC (P = 0.114). Significance differences between POC and SOC were found in CEA and CA125 level in the training and validation cohort (all P < 0.05).

A total of 396 radiomic features were extracted by AK software. After dimensionality reduction, which included ANOVA and KW, univariate logistic regression (14 features) removes the redundancy with correlation coefficient more than 0.90 (5 features) (Fig. 3a) and after the LASSO algorithm with λ value of 0.0245 and log (λ) value of 3.71, two significant radiomic features were identified [15, 16] (Fig. 3b). Two potential radiomic features including Percentile15 and Inverse Difference Moment_AllDirection_offset1_SD remained after dimension reduction with LASSO. The contribution of the selected features and their corresponding regression coefficients are shown in Fig. 3c. The two characteristics of the ROC curve of training and validation are shown in Fig. 3d–e.These features were presented in the rad-score calculated by using the following formula:

rad-score =  − 0.220 + 0.621*Percentile15 + 0.994*InverseDifferenceMoment_AllDirection_offset1_SD.

The Rad-score for each patient presented as a waterfall plot demonstrated significant differences between POC and SOC in both training (P = 0.001) and testing cohorts (P = 0.015) (Fig. 3f–g). The radiomics signature also showed a favorable predictive efficacy, with an AUC of 0.734 in the training cohort (95% CI 0.620–0.847, sensitivity = 51.5%, specificity = 83.7%) and 0.733 in the validation cohort (95% CI 0.549–0.917, sensitivity = 60%, specificity = 89.5%).

The rad-score, CEA, and CA125 were identifed as independent factors to distinguish POC and SOC by logistic regression. Then, we integrated the above factors into a prediction model. As shown in Fig. 4a–b, the combination model outperformed the radiomic features only model and clinical features only model with a greater AUC of 0.854 in training cohort and 0.751 in testing cohort. Finally, a radiomic-based nomogram for individualized differentiation was built by corporating the above radiomic features and 2 clinical factors (Fig. 5a). The total points accumulated by various variables correspond to the predicted probability for a patient. The calibration curve of the radiomics nomogram demonstrated good agreement in both training and validation cohorts. (Fig. 5b–c).