Diagnosis of Ovarian Neoplasms Using Nomogram in Combination With Ultrasound Image-Based Radiomics Signature and Clinical Factors

Objectives: To establish and validate a nomogram integrating radiomics signatures from ultrasound and clinical factors to discriminate between benign, borderline, and malignant serous ovarian tumors. Materials and methods: In this study, a total of 279 pathology-confirmed serous ovarian tumors collected from 265 patients between March 2013 and December 2016 were used. The training cohort was generated by randomly selecting 70% of each of the three types (benign, borderline, and malignant) of tumors, while the remaining 30% was included in the validation cohort. From the transabdominal ultrasound scanning of ovarian tumors, the radiomics features were extracted, and a score was calculated. The ability of radiomics to differentiate between the grades of ovarian tumors was tested by comparing benign vs borderline and malignant (task 1) and borderline vs malignant (task 2). These results were compared with the diagnostic performance and subjective assessment by junior and senior sonographers. Finally, a clinical-feature alone model and a combined clinical-radiomics (CCR) model were built using predictive nomograms for the two tasks. Receiver operating characteristic (ROC) analysis, calibration curve, and decision curve analysis (DCA) were performed to evaluate the model performance. Results: The US-based radiomics models performed satisfactorily in both the tasks, showing especially higher accuracy in the second task by successfully discriminating borderline and malignant ovarian serous tumors compared to the evaluations by senior sonographers (AUC = 0.789 for seniors and 0.877 for radiomics models in task one; AUC = 0.612 for senior and 0.839 for radiomics model in task 2). We showed that the CCR model, comprising CA125 level, lesion location, ascites, and radiomics signatures, performed the best (AUC = 0.937, 95%CI 0.905–0.969 in task 1, AUC = 0.924, 95%CI 0.876–0.971 in task 2) in the training as well as in the validation cohorts (AUC = 0.914, 95%CI 0.851–0.976 in task 1, AUC = 0.890, 95%CI 0.794–0.987 in task 2). The calibration curve and DCA analysis of the CCR model more accurately predicted the classification of the tumors than the clinical features alone. Conclusion: This study integrates novel radiomics signatures from ultrasound and clinical factors to create a nomogram to provide preoperative diagnostic information for differentiating between benign, borderline, and malignant ovarian serous tumors, thereby reducing unnecessary and risky biopsies and surgeries.


US image feature extraction
To normalize the different image specifications from various US scanners, image resampling and gray-level normalization were performed before quantitative feature extraction. All image data were resampled at a 1×1×1-mm voxel space size. The quantitative features were extracted from ROIs using an in-house software developed with MATLAB 2018B (MathWorksInc.).
A total of 885 radiomic features were drawn from each segmented lesion and can grouped as follows: (1) Morphologic features: four metrics, including area, largest diameter, length to width ratio and roundness, were calculated for the morphological description of the images. Area is the number of voxels in the tumor region extracted from US images multiplied by the dimension of voxels. Largest diameter is the voxels number of the long axis. Length to width ratio is the ratio of length to width. Roundness is defined as the ratio of the circumcircle radius to the inscribed circle radius of the lesion ROI.
(2) Gray-scale histograms features: three features were computed for each lesion according to the definitions of the gray-scale histogram: variance, skewness and kurtosis. Their definition can be found in literatures Rodenacker K (2003). (3) Texture features: in total, 40 texture features were extracted from the tumor regions of US images after wavelet transform. Table 1 presents the list of texture features used in this study. Detailed description and methodology employed to extract the texture features is available in M Valliè res (2015). (4) Wavelet features: wavelet transform effectively decouples textural information by decomposing the original image. In this study a discrete, one-level and undecimated two dimensional wavelet transform was applied to each US image, which decomposes the original image into 4 decompositions (LL, HL, LH and HH). For each decomposition we computed gray-scale histograms and the textural features as described in Table S1.

Task 1
Rad-score was calculated by summing the selected features weighted by their coefficients. The final formula of rad-score is: And we compared the Rad-scores from the training and validation cohort, respectively ( Figure S2 A-B). The cutoff value was: 0.633.

Task 2
Rad-score calculation formula: And we compared the Rad-scores from the training and validation cohort respectively ( Figure S2 C-D). The cutoff value was: 0.491. Figure S1: Feature selection using the least absolute shrinkage and selection operator (LASSO) binary logistic regression model for tow tasks. (A-B) Tuning parameter (Lambda) selection in the LASSO model used 10-fold cross-validation via minimum criteria for task 1 and 2, respectively. The gray line in the figure is the partial likelihood estimate corresponding to the optimal value of lambda. The optimal lambda value of 0.061 (task 2: 0.104) was chosen. (C-D) LASSO coefficient profiles of the features of task 1 and 2, respectively. A vertical line was plotted at the optimal lambda value, which resulted in 17 (task 1) and 22 (task 2) features with nonzero coefficients.  Figure S2: (A) The radscore from the training cohort for task 1. (B) The radscore from the validation cohort for task 1. (C) The radscore from the training cohort for task 2. (D) The radscore from the validation cohort for task 2.