The Value of Texture Analysis of Multi-parameter MRI Images in Distinguishing Benign and Malignant Lesions of O-RADS MRI 4 Lesions.

Ovarian tumors are the most common tumors of the female reproductive system and are not easily detected, especially in the early stages. Owing to their high morbidity and mortality, distinguishing between benign and malignant ovarian tumors is of great importance 1 . Currently, imaging remains the primary method of ovarian tumor detection. Ultrasound has the advantages of painlessness, simplicity of operation, and high repeatability in diagnosing ovarian tumors and is widely used. However, approximately 18%-31% of ovarian tumors cannot be accurately diagnosed using ultrasound examination 2 , 3 . Compared with ultrasound, magnetic resonance imaging (MRI) shows better anatomical resolution and soft-tissue contrast imaging 4 and is also a multi-parameter imaging technique. In this context, the ADNEX-MR score was introduced in 2013 as a risk assessment and adnexal lesion scoring system 4 , and its name was changed to ovarian-adnexal reporting and data system MRI (O-RADS MRI) score in 2020 5 . The O-RADS allows for the stratification of the malignancy risk of adnexal masses based on the lesion's composition, the signal intensity characteristics, and the solid tissue's enhancement pattern. The O-RADS MRI risk scoring system has promising diagnostic efficacy and reproducibility, but 10.8% to 12.5% of ovarian adnexal masses malignancies cannot be determined. In patient management dilemmas, O-RADS MRI 1-3 (normal to low-risk) and O-RADS MRI 5 (high-risk) usually do not lead to difficulties. Considering the wide range of malignant tumors in such lesions, the adnexal lesions with O-RADS MRI 4 (moderate risk) have been considered as the system's Achilles' heel. The positive predictive value (PPV) of malignant lesions with an O-RADS MRI risk score of 4 was only 50% 6 . In addition, when nondynamic multi-phase contrast-enhanced (CE) MRI is used to evaluate adnexal lesions, the PPV of O-RADS MRI 4 is still unclear and requires further study 4 , 5 , 7 . A lower PPV may lead to a considerable number of patients with benign adnexal tumors undergoing potentially unnecessary or excessive surgical intervention. Consequently, more criteria are required to sub-stratify O-RADS MRI 4 for better risk stratification. The O-RADS MRI risk-scoring system is still in the early stages of use, and further research is necessary to optimize and develop it. The O-RADS MRI score is principally based on MRI's morphological and functional manifestations and assesses tumors through macroscopic observation of lesion morphological characteristics and semi-quantitative evaluation. It has limited subjectivity and lacks objective quantitative data analysis. Texture analysis (TA) is a mathematical method that systematically and quantitatively analyzes image features to evaluate tissue gray-scale patterns, locations, and the relationship between pixels and voxels. It can quantify tumor heterogeneity and microstructure that the human visual system cannot detect 8 , 9 . Recently, texture analysis has been combined with various imaging methods to predict the pathological classification, staging, and postoperative prognosis of gynecological tumors 10 , 11 . Wang et al. 12 used a multimodal MRI texture analysis model to predict the prognosis of ovarian cancer and found that a model based on T2WI had the best performance. Ye et al. 13 explored the value of MRI-based whole-tumor texture analysis in distinguishing borderline and malignant epithelial ovarian tumors and found that the area under the receiver operating characteristic curve (AUC) of the combined model of texture features and clinical data was 0.962. There have been few reports on using TA to improve the O-RADS MRI scores to 4. Therefore, this study aimed to explore the value of TA based on multimodal MRI images to improve the O-RADS MRI score to 4.

This retrospective study was conducted between March 2015 and December 2022 at the Radiology Department of Minhang Hospital, affiliated with Fudan University, China. It included 147 patients with 155 adnexal masses confirmed surgically and pathologically. The Institutional Review Board of Minhang Hospital, affiliated with the Fudan University Review Board, approved this retrospective study, and informed consent was not required. The inclusion criteria were pelvic MRI enhancement examination within 4 weeks before surgery and a final O-RADS MRI lesion score of 4 points. The exclusion criteria were poor image quality, inability to be used for evaluation and analysis, missing or incomplete MRI images, incomplete clinical and pathological data, and preoperative radiation or chemotherapy. The final cohort comprised 52 patients with 57 adnexal masses. All MRI examinations were performed on a 3.0-T system (uMR780 3.0 T MRI; United Imaging Healthcare, Shanghai, China) and a 1.5-T system (EXCITE HD 1.5T MRI; GE Healthcare, Milwaukee, WI, USA) using a phased-array coil. Fasting was done for 4-6 hours before examination to limit artifacts caused by intestinal peristalsis. The scan range of the pelvis was from the iliac crest to the pubic symphysis and was adjusted to cover the entire size of the adnexal lesion. The contrast agent Magnevist was given at a dose of 0.2 mL per kilogram of body weight using a power injector at a rate of 2 mL/s, followed by 15 mL of normal saline to flush the tubing. Our institutional standard MRI protocol included the following sequences: axial and T1 fast spin-echo (FSE) weighted imaging (WI) with and without fat saturation; sagittal and axial T2 FSE WI; axial diffusion-weighted images (DWI) with b-values of 0, 800, 1000 s/mm 2 to obtain apparent diffusion coefficient (ADC) maps; dynamic T1weighted 3D gradient-echo with fat saturation in the axial and sagittal plane during contrast uptake and delayed post-contrast T1-weighted 3D gradient echo with fat saturation in the axial plane. The scanning sequences are presented in Table 1 . Two radiologists with 5 and 15 years of experience in female pelvic imaging analyzed all images independently. Both radiologists were blinded to the clinical and histological data, and retrospectively classified the adnexal masses according to the O-RADS MRI scoring system published by Thomassin et al. in January 2020 5 . If there is a disagreement between the two readers a consensus can be reached through discussion. According to previously published studies, the following MRI characteristics were analyzed for each adnexal mass 14 - 17 : distribution location (unilateral, bilateral), shape (regular, irregular), boundary (clear, unclear), lesion diameter (maximum cross-sectional diameter of the lesion), lesion composition (cystic, cystic solid, solid), T1WI image signal strength, T2WI lipid-pressure image signal intensity, DWI signal strength, enhancement mode, ascites (with or without), lymph nodes (with or without). The signal intensity of the endometrium was used as the reference standard for the signal intensity of lesion DWI, and the degree of enhancement of the myometrium was used as the reference standard for the degree of enhancement of the lesion. Patients with a score of 4 on the O-RADS non-dynamic enhanced MRI were enrolled. Adnexal lesion segmentation was performed using ITK-SNAP software ( http://www.itk-snap.org ). Regions of interest (ROI) were manually segmented independently by two radiologists with 5 and 15 years of experience in MRI image evaluation, respectively, layer by layer for the entire lesion. Both radiologists reached a consensus in cases with obscured tumor margins by performing additional image analyses. An Artificial Intelligence Kit (A. K., GE Healthcare) was used to analyze the volume of interest (VOI) and extract lesion texture features. A total of 3474 features were obtained, including shape features, first-order statistics, gray-level co-occurrence matrix (GLRLM, gray-level run length matrix (GLSZM, gray-level size zone matrix (GLDM, gray-level dependence matrix (GLDM), and Laplace of Gaussian transformed features (LoG). PyRadiomics ( http://www.radiomics.io/pyradiomics.html ) was used to extract 704 wavelet transform features. Detailed information on the textural features is presented in Table 2 . The intraclass correlation coefficient (ICC) of the textural features of 30 cases was randomly selected to evaluate the intra- and inter-observer repeatability of lesion segmentation. Radiologist 1 manually plotted the VOI twice within two weeks for intra-observer ICC assessment. Radiologist 2 also outlined the VOI, and the extracted texture features were used to evaluate the ICC between observers. ICC > 0.75 showed good consistency. The Mann-Whitney U test was used for univariate analysis. The texture features were retained when the intra-and inter-observer ICCs were greater than 0.75 and p < 0.1. Subsequently, the mRMR method was used to select the feature subset, and features with minimum redundancy and maximum correlation were retained. Subsequently, four prediction models of T2WI, ADC, T1WI, and three combined models were constructed using the random forest (RF) algorithm. Ten-fold cross-validation was used to verify the model's predictive performance, and ROC curve analysis was used to evaluate the model's performance. The areas under the curve AUC, accuracy, sensitivity, specificity, PPV, and NPV were also recorded. After surgical resection, all specimens were formalin-fixed, dehydrated, paraffin-embedded, sectioned, and stained with hematoxylin-eosin staining (HE staining). Two senior pathologists observed the sections independently. If HE staining cannot be used for diagnosis, immunohistochemistry can be used to further determine the pathological type. All statistical analyses were conducted using SPSS statistical software (version 26.0; IBM Corp., Armonk, NY, USA) and R software (version 4.2.0; http://www.r-project.org ). The Kolmogorov-Smirnov test determined whether the measurement data conformed to a normal distribution. The measurement data obeying the normal distribution was tested by independent sample T-test, expressed as mean ± standard deviation (x ± SD). Measurement data that did not follow a normal distribution were tested using the rank sum test and expressed as median and quartile spacing [M (P25, P75)]. The chi-square or Fisher's exact test was used to compare intergroup differences in categorical variables. The intragroup correlation coefficient was used for the observer consistency test. ICC assessed the consistency between observers using the following criteria: excellent 0.75-1.00; good 0.50-0.75; acceptable 0.25-0.50; very poor 0-0.25. p < 0.05 was considered significant.

The patients' baseline clinical and pathological data are shown in Tables 3 and 4 . This study enrolled 52 patients with 57 lesions (five cases of bilateral lesions) aged 17-86 years, with an average age of 51.5 years. Pathological results: 26 cases (45.6%) were benign lesions (8 cases of serous cystadenoma, 7 cases of mucinous cystadenoma, 3 cases of goiter, 3 cases of ovarian fibroma, 1 case of Brenner tumor, 2 cases of theca cell tumor, 2 cases of ovarian salpingitis mass), 31 cases (54.4 %) were malignant (3 cases of borderline serous tumor, 5 cases of borderline mucinous tumor, 4 cases of serous adenocarcinoma, 1 case of mucinous adenocarcinoma, 3 cases of clear cell carcinoma, 5 cases of endometrioid carcinoma, 2 cases of granular cell tumor, 1 case of malignant teratoma, and 7 cases of metastatic carcinoma). Among the 57 patients, 26 were postmenopausal, 31 were premenopausal, 15 were postmenopausal in patients with benign tumors, and 11 were postmenopausal in patients with malignant tumors. Among the 57 patients, 34 had elevated serum carbohydrate antigen 125(CA125) levels, one had elevated alpha-fetoprotein (AFP) levels, and five had elevated serum carcinoembryonic antigen (CEA) levels. There were no significant differences in age, serum (CA125) level, alpha-fetoprotein (AFP) level, carcinoembryonic antigen (CEA) level, and menopause between patients with benign and malignant tumors ( p > 0.05). A total of 3474 texture features were extracted from ADC, T2WI, and CE_T1WI images. Significant texture features were selected by single-factor analysis, and the ADC, T2WI, T1WI enhanced, and combined models were established using the mRMR and RF algorithms. Each model contained ten texture features. The ten texture features in the ADC model include two first-order features, two LoG features, and six wavelet features. The ten texture features in the T2WI model included three LoG features and seven wavelet features. Ten texture features in the T1WI enhancement model included five first-order features, one LoG feature, and four wavelet features. Ten texture features in the combined model included three first-order features, one LoG feature, and six wavelet features. The ICC of the texture features of the model was greater than 0.75, and the results of the single-factor analysis of the selected features of the four models are shown in Table 5 . The AUC comparison of the 10 independent predictive factors selected by the four models is shown in Figure 1 , and the important components of the 10 independent predictive factors in the four models are shown in Figure 2 , among which the wavelet_LHH_glcm_MCC.ADC features account for the largest proportion of the selected ten features in the combined model. The correlation between the expression of the 10 independent predictive factors in the heat map of the four models and malignancy is shown in Figure 3 . The ROC curves for the four models are shown in Figure 4 . The AUC of the ADC, T2WI, and CE_T1WI models were 0.749 (95 % CI: 0.621-0.876), 0.671 (95 % CI: 0.524-0.818), and 0.786 (95 % CI: 0.662-0.909), respectively. The Delong test showed that there was no difference in AUC among the three model groups (T2WI vs CE_T1WI model, p = 0.101; T2WI vs. ADC model, p = 0.347; CE_T1WI vs ADC model, p = 0.597). The AUC of the combined model was 0.860 (95% CI: 0.76-0.959). Delong tests showed that the AUC between the T2WI, ADC, and combined models was statistically significant (T2WI and combined model, p = 0.004; ADC vs. combined model, p = 0.038; CE_T1WI model vs combined model p = 0.070). The AUC of the combined model was significantly higher than those of the other three groups. A ten-fold cross-validation of the combined model is shown in Figure 5 . The combined model exhibited the best diagnostic performance, with an accuracy of 0.759. After ten-fold cross-validation, it performed well, with an accuracy of 0.713. Table 6 presents the diagnostic performance of the four models. The combined model had the best performance, with accuracy, sensitivity, specificity, PPV, and NPV of 75.9%, 77.8%, 74.1%, 72.4%, and 79.3%, respectively.

This study evaluated the diagnostic performance of TA in improving the diagnostic performance of O-RADS 4. Our results showed that the texture features from the T2WI, ADC, and CE_T1WI images significantly differed between benign and malignant O-RADS MRI 4. The combined model based on textural features showed promising efficiency in distinguishing between benign and malignant O-RADS 4 lesions (AUC = 0.86). The O-RADS MRI risk scoring system is based on the MRI scoring system for adnexal lesions (ADNEX MR scoring system) developed by Thomassin et al. in 2013 4 . It was developed and published by a multidisciplinary, international expert committee under the guidance of the American College of Radiology in 2020 as the most comprehensive standard for MRI evaluation of ovarian and adnexal lesions. A meta-analysis by Qing Zhang et al. 18 found that O-RADS US and O-RADS MRI have high sensitivity for ovarian or adnexal malignancies. However, O-RADS MRI provides higher specificity (90%). A meta-analysis involving 4,012 ovarian adnexal lesions by Stefania Rizzo et al. 19 found that the total malignant probability of O-RADS MRI 4 lesions was 60% (95% CI, 52%-67%). This study quantified the texture features of whole lesions of ovarian tumors based on multi-parameter MRI TA from magnetic resonance ADC maps, T2WI, and CE_T1WI. Our results indicate that the combined model based on the ADC map, T2WI, and T1WI enhanced sequences can improve the diagnostic efficacy between benign and malignant lesions with an O-RADS MRI score of 4. Quantified texture features compensate for empirical diagnoses based on conventional MRI morphology deficiencies. The MRI-based texture feature model provides a new method for noninvasively improving the diagnostic efficacy of O-RADS 4 before surgery, which helps evaluate preoperative treatment strategies. TA is a new imaging technology that can extract large amounts of data from biomedical images and be studied using TA tools. TA of the MRI images was performed by analyzing the gray tone changes between the image voxels. This method captures spatial and intensity information to identify pathological changes and microstructural heterogeneity within tumors. Because of the heterogeneity of ovarian tumors, texture features based on the entire tumor contain more spatial information and higher sensitivity and specificity than traditional methods based on pre-selected regions of interest. This study used 3D whole-lesion ROI delineation to avoid the limitations of selecting individual layers corresponding to partial lesion areas and maximizing their diversity 20 . 3D TA is a method that utilizes all data dimensions and contains rich information about the internal structure of related objects. Increasing the number of points required for feature calculation to provide a more complete description of the lesion can improve the accuracy of lesion heterogeneity characterization and reduce sampling errors 21 , 22 . There were subtle grey differences between the tumor and its surrounding tissues, but the difference was not obvious; areas with more low-frequency and high-frequency information appeared randomly. The wavelet transform relies mainly on multi-resolution analysis to decompose an image into components of different frequencies, followed by data correction. The multi-resolution analysis characteristics of the wavelet transform make it widely applicable to the texture features of non-stationary signals, such as medical images 23 . In this study, among the ten features selected by the combined model, there were six high-order wavelet transform features. TA was performed in the spatial and frequency domains to provide changes in spatial resolution and represent texture on the most appropriate scale. Wavelet transform features enhance the details of an image, which is of great significance in this study. Among the 10 features the combined model selected, three come from GLCM, where the feature wavelet_LHH_glcm_MCC.ADC_DWI_b = 800 accounts for the largest proportion among the selected ten features. The relationship between two pixels can influence the GLCM and calculate the number of occurrences of all possible combined gray values in a specific direction and the distance between them. The GLCM is calculated for multiple directions and distances, retains only the required solutions, and presents the best features. It is one of the most important and in-depth TA methods and many applications. It remains the benchmark method for most TA research 24 - 26 . Fathi Kazerooni et al. 27 showed that GLCM texture parameters based on magnetic resonance ADC maps have value in the differential diagnosis between benign and malignant solid ovarian tumors. Combined with our findings, we assume that GLCM features may provide unique information regarding ovarian tumors. Since the introduction of radiomics in 2012, TA has been widely used to study ovarian tumors. RA et al. 28 established a prediction model based on the radiomic TA of MRI images of liquid components in ovarian cysts. The prediction model's sensitivity, specificity, and AUC for identifying malignant lesions were 84.62%, 80%, and 0.841, respectively. Zhang et al. 29 established a radiomics model based on the texture features of multiple sequences of MR images (T1WI, T2WI, T2WI_fs, DWI, and T1WI multi-phase enhancement), which had higher diagnostic accuracy than radiologists in differentiating benign and malignant ovarian tumors (90.6% and 83.5%, respectively). The model could distinguish type I and type II epithelial ovarian cancers with an accuracy of 83% and an AUC of 0.85, and its diagnostic performance was better than when the features were applied separately. A few researchers have attempted to combine MRI radiomics to improve the effectiveness of O-RADS MRI. Hottat et al. 30 applied DWI quantitative analysis, including ROI-ADC and whole lesion-ADC histogram measurements combined with O-RADS MRI 4, and the overall performance of O-RADS was improved. However, no studies have focused on applying MRI-based textural features to improve the performance of O-RADS MRI. Multi-parameter TA based on MRI has been widely used to improve the magnetic resonance stratification management score of breast, prostate, and other tumors 31 - 33 . This study focused on the multi-parameter MRI texture features of magnetic resonance ADC maps, T2WI, and CE_T1WI to improve the effectiveness of O-RADS MRI, which showed good clinical practicability. Although the TA model based on multi-parameter MRI had good accuracy, this study had some limitations. First, this was a retrospective study, and selection bias may exist. Second, the sample size was small, therefore, a multi-center study with an expanded sample size is necessary to verify the applicability of the results. In addition, although this study explored the preliminary results of MRI image texture analysis to improve the diagnostic efficacy of O-RADS MRI with a score of 4 for benign and malignant tumors, further studies are necessary to confirm the pathological basis. In summary, the model based on the whole lesion of the ovarian tumor texture of multi-parameter MRI helps improve the diagnostic efficiency of the O-RADS MRI score of 4 for benign and malignant tumors. This approach offers a promising solution to the existing challenges in differentiating O-RADS MRI 4 lesions, while also addressing the variability in clinical expertise among imaging diagnostic physicians—particularly junior practitioners who predominantly rely on conventional MP-MRI morphological features for evaluation. Moving forward, we aim to undertake large-scale, multicenter prospective cohort studies to further assess the practical utility of O-RADS MRI scores in real-world clinical settings. These efforts aim to provide an objective and accurate diagnosis and treatment basis for clinical practice, assists in developing personalized diagnosis and treatment plans, and improves patient prognosis and survival rates.