Radiomics analysis based on T2-weighed imaging and T2 mapping for staging endometrial fibrosis.

This prospective study was conducted in accordance with the Declaration of Helsinki (as revised in 2013) and was approved by the Ethics Committee of the Nanjing Drum Tower Hospital (approval no. 2019‑051‑04). Written informed consent was obtained from all participants. According to the American Fertility Society (AFS) scoring criteria, patients with severe IUA display more fibrous scar tissue than patients with mild-moderate IUA 25 . Therefore, Mild-moderate IUA was defined as MMEF, while severe IUA was considered as SEF 19 . Representative hysteroscopic images were shown in Fig.  1 . A total of 128 SEF patients, 52 MMEF patients and 103 healthy controls (HC) from September 2018 to August 2023, respectively. For continuity, this study included some patients and controls that overlapped with our previous study 19 . Fig. 1 Representative hysteroscopic images of normal endometrium, mild-moderate endometrial fibrosis (MMEF), and severe endometrial fibrosis (SEF). ( a ) smooth reddish endometrium in healthy women. ( b ) fibrotic adhesions with some normal endometrium in MMEF patients. ( c ) unsmooth endometrium with dense fibrotic adhesions in SEF patients. Representative hysteroscopic images of normal endometrium, mild-moderate endometrial fibrosis (MMEF), and severe endometrial fibrosis (SEF). ( a ) smooth reddish endometrium in healthy women. ( b ) fibrotic adhesions with some normal endometrium in MMEF patients. ( c ) unsmooth endometrium with dense fibrotic adhesions in SEF patients. The inclusion criteria for patients were as follows: (1) aged 20–42 years with normal ovarian function; (2) clinical diagnosis of infertility (no pregnancy after 12 months of regular unprotected intercourse) 26 and endometrial fibrosis confirmed by hysteroscopy; (3) history of dilation and curettage (D&C). The exclusion criteria were as follows: (1) severe uterine diseases such as severe adenomyosis and endometrial hyperplasia, or severe malformations of uterus according to the ESHRE/ESGE consensus 27 ; (2) MRI contraindications. The inclusion criteria for HC were as follows: (1) aged 20–42 years; (2) regular menstrual cycles and normal volumes. The exclusion criteria for HC were as follows: (1) history of D&C or severe uterine diseases; (2) MRI contraindications. All patients were diagnosed with endometrial fibrosis via hysteroscopy either prior to the MRI scans, with an interval of approximately 1–2 months to minimize endometrial acute changes caused by hysteroscopy, or within 2 days following the MRI scans. MRI examinations were performed with a 3.0-T MRI scanner (Ingenia, Philips Medical Systems, Best, The Netherlands) using a 16-channel phased-array body coil in the head-first and supine position. To avoid the physiological interference from the menstrual cycle, all participants underwent MRI scans at the late proliferative phase (with a dominant follicle confirmed by ultrasonography). Each participant was asked to empty the bladder and fast for 2 h to reduce motion artifacts. The MRI sequences were acquired as follows: (I) sagittal T2-weighted imaging (T2WI) using turbo spin echo (TSE) [TR = 3691 ms, TE = 100 ms, TSE factor = 25, flip angle = 90°, matrix = 200 × 167, voxel size = 0.6 × 0.7 × 3 mm, FOV = 120 mm×120 mm, slice thickness = 3 mm, slice gap = 0.3 mm, slices = 18, scan duration = 3 min 56 s]; (II) sagittal T2 mapping using TSE [TR = 1773 ms, TEs = 17, 34, 51, 68, and 85 ms, TSE factor = 5, flip angle = 90°, matrix = 124 × 108, voxel size = 1.6 × 1.75 × 3 mm, FOV = 200 mm×200 mm, slice thickness = 3 mm, slice gap = 0.3 mm, slices = 18, scan duration = 57 s]. T2 maps were automatically generated using the embedded software in the MRI scanner after data acquisition. Image quality was assessed by two radiologists (with 5 and 3 years of experience, respectively) in consensus. Images exhibiting motion artifacts or other artifacts compromising endometrium definition were deemed insufficient in quality, which was excluded from subsequent analysis. Two radiologists jointly used 3D Slicer (version 5.5.0, http://www.slicer.org ) to semiautomatically segment the whole endometrium of corpus uteri on sagittal T2WI images and T2 maps, excluding visible hemorrhagic and cystic areas. All disagreements were confirmed by the third senior radiologist with 9 years of experience. The radiologists were blinded to the clinical information of all participants. And 30 cases from all participants were randomly selected to assess the stability of radiomic features on sagittal T2WI images and T2 maps, respectively. The first radiologist performed image segmentation twice with an interval of one month to evaluate the intraobserver reproducibility. The second radiologist independently delineated the segmentation to evaluate the interobserver reproducibility. In the pre-processing step, the N4 bias field correction algorithm was applied only to the T2WI images to address image inhomogeneity. Prior to radiomic feature extraction, z-score normalization of MRI signal intensities was performed exclusively on T2WI to reduce inter-subject variability. No bias field correction or normalization was applied to the T2 maps to preserve their intrinsic quantitative values. For both T2WI images and T2 maps, gray-level discretization with a fixed bin width of 25 and voxel size resampling to 1 × 1 × 1 mm were performed using PyRadiomics (version 3.0.1; https://github.com/Radiomics/PyRadiomics ) to reduce variability associated with differing voxel sizes and enhance the reproducibility of extracted features. Subsequently, feature extraction was performed on the whole endometrium of the corpus uteri (3D volume). Radiomic feature extraction was performed using the PyRadiomics Image Biomarker Standardization Initiative (IBSI) consensus python library. The extracted features include shape, first-order histogram, gray level co-occurrence matrix (GLCM), gray level run length matrix (GLRLM), gray level size zone matrix (GLSZM), gray level dependence matrix (GLDM), and neighboring gray tone difference matrix (NGTDM). In addition, 6 image filters were applied to the original images to generate derived images, including Exponential, Gradient (kernel sizes: 1.0 mm, 3.0 mm, and 5.0 mm), Logarithm, Square, SquareRoot, and Wavelet filters (LLL, LLH, LHL, HLL, LHH, HLH, HHL, HHH). In this study, the radiomic features with ICCs (both intraobserver ICC and interobserver ICC) greater than 0.8 and without high correlation between each other (Spearman ρ  < 0.95) were regarded as reliable and retained for dimensionality reduction 28 , 29 . Finally, the recursive feature elimination (RFE) based on random forest (RF) by 10-fold cross-validation was applied to select important features. All Participants were randomly assigned to training set and test set at a ratio of 7:3. The synthetic minority over-sampling technique (SMOTE) was applied to balance the minority class of MMEF patients in the training set for the 3-class classification model 30 . The training set was used to train logistic regression (LR), support vector machine (SVM), and decision tree (DT) classifiers for constructing various classification models: T2WI model, T2 mapping model, and merged model (both T2WI images and T2 maps), respectively. Fivefold cross-validation was applied to tune hyperparameters for better model performance. Finally, we compared the performance of the radiomics models using sensitivity, specificity, precision, F1-score, and area under the receiver operating characteristic curve (AUC) to identify the optimal radiomics model. The overall workflow of the radiomics model development is shown in Fig.  2 . Fig. 2 Radiomics flow diagram of the study. VOI, volume of interest; SMOTE, synthetic minority over-sampling technique; RF-RFE, random forest with recursive feature elimination; LR, logistic regression; SVM, support vector machine; DT, decision tree; ROC, receiver operating characteristic curve; HC, healthy controls; EF, endometrial fibrosis; MMEF, mild-moderate endometrial fibrosis; SEF, severe endometrial fibrosis. Radiomics flow diagram of the study. VOI, volume of interest; SMOTE, synthetic minority over-sampling technique; RF-RFE, random forest with recursive feature elimination; LR, logistic regression; SVM, support vector machine; DT, decision tree; ROC, receiver operating characteristic curve; HC, healthy controls; EF, endometrial fibrosis; MMEF, mild-moderate endometrial fibrosis; SEF, severe endometrial fibrosis. T2WI images of the uterus were transferred to ImageJ software (version 1.54i) to measure endometrial thickness (ET) of the corpus uteri by two radiologists independently. ET was measured at the thickest part of the endometrium. The final ET was calculated by the average value measured by the two radiologists. The endometrial mean T2 value (Mean T2) was extracted as one of the histogram features from the VOI segmentation on the T2 maps. Finally, the ET+T2 model was constructed based on ET and mean T2 value using logistic regression. Continuous variables with normal distribution were presented as mean ± standard deviation, while continuous variables of no normal distribution were recorded as median (interquartile range [IQR]). The chi-square or Fisher’s exact tests were used for diagnosis distribution, while independent t-test or Mann-Whitney U test were used for age and imaging features between training set and test set. The diagnostic performance of the model was assessed with receiver operating characteristic (ROC) curve analysis and AUC. The comprehensive performance evaluation of the multiclass model was performed using macro-average and micro-average methods 31 . Statistical analyses were performed using R software (version 4.0.4; https://www.r-project.org/ ). A two-sided p -value less than 0.05 was indicative of a significant difference.

The flow diagram of the study population enrolled is shown in Fig.  3 . A total of 100 HCs, 50 patients with MMEF, and 120 patients with SEF were ultimately included. Representative T2WI images and T2 maps in HC, MMEF patient and SEF patient are shown in Fig.  4 . There were no significant differences in baseline characteristics between the training set and test set (Table  1 ). The median ET was 10.8 mm (IQR: 9.2–12.2 mm) in HCs, 8.4 mm (IQR: 7.5–9.6 mm) in MMEF patients, and 6.6 mm (IQR: 5.5–7.6 mm) in SEF patients, respectively. The median endometrial mean T2 value was 218 msec (IQR: 186–255 msec) in HCs, 145 msec (IQR: 129–168 msec) in MMEF patients, and 132 msec (IQR: 115–158 msec) in SEF patients, respectively. The ET and mean T2 value differed significantly between the three groups (both p  < 0.001). Fig. 3 Flow diagram of the study population. HC, healthy controls; MMEF, mild-moderate endometrial fibrosis; SEF, severe endometrial fibrosis; D&C, dilation and curettage. Flow diagram of the study population. HC, healthy controls; MMEF, mild-moderate endometrial fibrosis; SEF, severe endometrial fibrosis; D&C, dilation and curettage. Fig. 4 Midsagittal T2WI images and T2 maps in a healthy control (HC) ( a , b ), a patient with mild-moderate endometrial fibrosis (MMEF) ( c , d ), and a patient with severe endometrial fibrosis (SEF) ( e , f ). T2WI images ( a , c , e ) show that the endometrial thickness (ET) of corpus uteri were 12.1 mm in the HC, 8.0 mm in the MMEF patient and 5.5 mm in the SEF patient. T2 maps ( b , d , f ) show that endometrial mean T2 was 241 msec in the HC, 147 msec in the MMEF patient, and 115 msec in the SEF patient. Midsagittal T2WI images and T2 maps in a healthy control (HC) ( a , b ), a patient with mild-moderate endometrial fibrosis (MMEF) ( c , d ), and a patient with severe endometrial fibrosis (SEF) ( e , f ). T2WI images ( a , c , e ) show that the endometrial thickness (ET) of corpus uteri were 12.1 mm in the HC, 8.0 mm in the MMEF patient and 5.5 mm in the SEF patient. T2 maps ( b , d , f ) show that endometrial mean T2 was 241 msec in the HC, 147 msec in the MMEF patient, and 115 msec in the SEF patient. Table 1 Baseline characteristics in the training set and test set. Characteristic Training set ( n  = 189) Test set ( n  = 81) p -value Diagnosis (%) 1.000 HC 70 (37.0) 30 (37.0) MMEF 35 (18.5) 15 (18.5) SEF 84 (44.5) 36(44.5) Age (years) 31 (26, 34) 32 (27, 36) 0.204 ET (mm) 8.1 (6.5, 10.4) 8.3 (6.4, 10.0) 0.898 Mean T2 (msec) 159 (129, 216) 154 (122, 209) 0.384 Data are presented as number (percentage) or median (interquartile range). HC, healthy controls; MMEF, mild-moderate endometrial fibrosis; SEF, severe endometrial fibrosis; ET, endometrial thickness; Mean T2, endometrial mean T2 value. Baseline characteristics in the training set and test set. Data are presented as number (percentage) or median (interquartile range). HC, healthy controls; MMEF, mild-moderate endometrial fibrosis; SEF, severe endometrial fibrosis; ET, endometrial thickness; Mean T2, endometrial mean T2 value. 1409 radiomic features were extracted from T2WI images and T2 maps, respectively. Among them, 184 T2WI and 172 T2 maps radiomic features with ICCs lower than 0.8 were excluded. After hierarchical clustering, 355, 274, and 615 features were obtained for the T2WI, T2 maps, and merged models, respectively. The SMOTE was used to balance the minority class in the training set to obtain a relatively balanced dataset (HC: 70 women; MMEF: 70 patients; SEF: 84 patients). In the final feature selection using the recursive feature elimination, 5, 9, and 11 radiomic features were ultimately included in the T2WI, T2 maps, and merged models, respectively. The selected features of each single and merged models are available in the supplementary material (Supplementary Table S1 ). Selected features in the merged models are shown in Fig.  5 . Fig. 5 Feature importance selection based on recursive feature elimination (RFE) in merged radiomics model. Feature importance selection based on recursive feature elimination (RFE) in merged radiomics model. Table  2 presents the overall performance of nine radiomics model in the test set. Amongst them, the merged model constructed by LR showed the best classification performance with the macro and micro average AUC of 0.897 and 0.898, sensitivity of 0.744 and 0.873, specificity of 0.880 and 0.816, precision of 0.738 and 0.873, FI-score of 0.740 and 0.873, respectively. The detailed results of other radiomics model in the training set and test set are shown in Supplementary Table S2 – S4 . Table 2 Overall performance of different models in the test set for the selection of the optimal model. Model AUC (macro/micro) Sensitivity (macro/micro) Specificity (macro/micro) Precision (macro/micro) F1-score (macro/micro) LR_T2WI 0.873/0.883 0.705/0.828 0.864/0.771 0.698/0.840 0.691/0.834 LR_ T2 mapping 0.813/0.833 0.643/0.831 0.839/0.702 0.661/0.794 0.639/0.812 LR_ merged 0.897/0.898 0.744/0.873 0.880/0.816 0.738/0.873 0.740/0.873 SVM_T2WI 0.885/0.879 0.665/0.794 0.848/0.714 0.697/0.794 0.650/0.794 SVM_T2 mapping 0.780/0.815 0.522/0.779 0.789/0.591 0.562/0.696 0.530/0.736 SVM_ merged 0.886/0.863 0.704/0.872 0.861/0.760 0.701/0.828 0.691/0.852 DT_T2WI 0.836/ 0.863 0.650/0.823 0.847/0.723 0.655/0.812 0.649/0.818 DT_T2 mapping 0.737/0.769 0.533/ 0.723 0.789/0.644 0.580/0.740 0.543/0.734 DT_ merged 0.867/0.872 0.670/0.838 0.857/0.720 0.679/0.802 0.666/0.820 Bold values indicate the highest AUC, sensitivity, specificity, precision, and F1-score in each column, corresponding to the optimal model. LR, logistic regression; SVM, support vector machine; DT, decision tree; AUC, area under the receiver operating characteristic curve. Overall performance of different models in the test set for the selection of the optimal model. Bold values indicate the highest AUC, sensitivity, specificity, precision, and F1-score in each column, corresponding to the optimal model. LR, logistic regression; SVM, support vector machine; DT, decision tree; AUC, area under the receiver operating characteristic curve. As shown in Table  3 , the diagnostic efficiency of the LR-based merged model was better than that of the ET+T2 model, with overall accuracy of 0.826 (95%CI 0.769–0.873) and 0.705 (95%CI 0.641–0.764) in the training set, and 0.765 (95%CI 0.658–0.852) and 0.593 (95%CI 0.478–0.700) in the test set, respectively. The confusion matrices for the ET+T2 and merged models showed that the LR-based merged model had the highest percentage of correct predictions in the training and test sets (Fig.  6 ). The ROC curves showed that the LR-based merged model achieved superior classification than that of the ET+T2 model, with AUCs of 0.930 vs. 0.779, 0.860 vs. 0.715, and 0.904 vs. 0.790 for distinguishing HC, MMEF, and SEF, respectively (Fig.  7 ). Table 3 The performance of the ET+T2 model and the optimal radiomic model in the training set and test set CI, confidence interval; AUC, area under the receiver operating characteristic curve; HC, healthy controls; MMEF, mild-moderate endometrial. Model Training set Test set AUC (95% CI) Sensitivity Specificity Precision F1-score AUC (95% CI) Sensitivity Specificity Precision F1-score ET+T2 model HC 0.882 (0.832–0.932) 0.700 0.903 0.766 0.731 0.779 (0.662–0.896) 0.600 0.882 0.750 0.667 MMEF 0.782 (0.719–0.844) 0.557 0.818 0.582 0.569 0.753 (0.620–0.886) 0.467 0.742 0.292 0.359 SEF 0.904 (0.864–0.943) 0.833 0.836 0.753 0.791 0.832 (0.746-0,919) 0.639 0.778 0.697 0.667 Overall accuracy 0.705 (0.641–0.764) 0.593 (0.478-0.700) Macro/Micro 0.856/0.864 0.697/0.805 0.852/0.650 0.700/0.656 0.653/0.764 0.788/0.786 0.569/0.667 0.801/0.833 0.580/0.667 0.564/0.667 LR_ merged model HC 0.980 (0.965–0.994) 0.871 0.981 0.953 0.910 0.930 (0.879–0.981) 0.733 0.902 0.815 0.772 MMEF 0.893 (0.852–0.934) 0.743 0.877 0.732 0.738 0.860 (0.748–0.971) 0.667 0.894 0.588 0.625 SEF 0.946 (0.919–0.973) 0.857 0.878 0.809 0.832 0.904 (0.835–0.973) 0.833 0.844 0.811 0.822 Overall accuracy 0.826 (0.769–0.873) 0.765 (0.658–0.852) Macro/Micro 0.939/0.949 0.824/0.883 0.911/0.905 0.824/0.917 0.827/0.899 0.897/0.898 0.744/0.873 0.880/0.816 0.738/0.873 0.740/0.873 ET, endometrial fibrosis; HC, healthy controls; MMEF, mild-moderate endometrial fibrosis; SEF, severe endometrial fibrosis; LR, logistic regression. The performance of the ET+T2 model and the optimal radiomic model in the training set and test set CI, confidence interval; AUC, area under the receiver operating characteristic curve; HC, healthy controls; MMEF, mild-moderate endometrial. ET, endometrial fibrosis; HC, healthy controls; MMEF, mild-moderate endometrial fibrosis; SEF, severe endometrial fibrosis; LR, logistic regression. Fig. 6 Confusion matrices of different models. ( a ) Confusion matrix for the ET+T2 model in the training set. ( b ) Confusion matrix for the ET+T2 model in the test set. ( c ) Confusion matrix for the LR-based merged radiomics model in the training set. ( d ) Confusion matrix for the LR-based merged radiomics model in the test set. The values in the confusion matrices represent the number and percentage of correctly predicted samples for each actual class. Confusion matrices of different models. ( a ) Confusion matrix for the ET+T2 model in the training set. ( b ) Confusion matrix for the ET+T2 model in the test set. ( c ) Confusion matrix for the LR-based merged radiomics model in the training set. ( d ) Confusion matrix for the LR-based merged radiomics model in the test set. The values in the confusion matrices represent the number and percentage of correctly predicted samples for each actual class. Fig. 7 ROC curves analysis for different models. ( a , b ) ROC curves of the ET+T2 model in the training and test sets. ( c , d ) ROC curves of the LR-based merged radiomics model in the training and test sets. The ROC curves of macro-average and micro-average demonstrated that the overall discriminability of the LR-based merged radiomics model was superior to that of the ET+T2 model. ROC, receiver operating characteristic curve; AUC, area under the curve; HC, healthy controls; MMEF, mild-moderate endometrial fibrosis; SEF, severe endometrial fibrosis. ROC curves analysis for different models. ( a , b ) ROC curves of the ET+T2 model in the training and test sets. ( c , d ) ROC curves of the LR-based merged radiomics model in the training and test sets. The ROC curves of macro-average and micro-average demonstrated that the overall discriminability of the LR-based merged radiomics model was superior to that of the ET+T2 model. ROC, receiver operating characteristic curve; AUC, area under the curve; HC, healthy controls; MMEF, mild-moderate endometrial fibrosis; SEF, severe endometrial fibrosis.

This prospective study demonstrates that radiomics analysis based on T2WI and T2 mapping held the potential for the rapid and non-invasive staging of endometrial fibrosis. Our results revealed that ET and endometrial mean T2 value were helpful to distinguish among HC, MMEF, and SEF, but lack accuracy. Among the nine radiomics models, LR-based merged model improved relative performance over ET+T2. This study found that ET and mean T2 value of EF patients were lower than that of HCs, which was similar to the previous study 19 , 32 . This observation may be explained by the presence of fibrotic endometrium, scar tissue, and impaired endometrial vascularity, which contribute to endometrial thinning and decreased T2 values 20 . Simple ET is commonly used to assess the morphological changes in endometrial fibrosis, while the T2 value reflects the water content within the endometrial tissue. However, ET and endometrial T2 values cannot accurately capture the microstructural heterogeneity of fibrotic endometrium. In this study, ET+T2 model had the ability to distinguish between SEF, MMEF and HC, but it only exhibited lower overall accuracy of 0.593. Additionally, a recent study by Zhou et al. 33 demonstrated that diffusion-weighted imaging (DWI) could effectively stage endometrial fibrosis. However, this approach still relies on voxel-averaged metrics and may not fully represent tissue complexity. Therefore, the performance of conventional MRI parameters such as ET, T2 value, and DWI in accurately staging endometrial fibrosis requires further improvement. Radiomics enhances endometrial characterization by extracting high-dimensional features and subtle changes beyond visual assessment, potentially improving the performance of radiomics models 34 . In this study, three multiclass classifiers were used to explore the best radiomics models for directly staging endometrial fibrosis. Among the nine models, the merged model constructed by LR outperformed other models for differentiating between SEF, MMEF and HC, and showed satisfactory performance. To overcome class imbalances when training the classifiers, we used SMOTE algorithm. This method effectively reduced the training bias and improved overall classification performance 35 , 36 . Among all the features in the merged model, there were 1 feature from T2 mapping and 10 features from T2WI images, suggesting that T2WI images have a greater impact on staging endometrial fibrosis. Moreover, most of the features selected for this model are wavelet-transformed features. Wavelet transformation is a higher-order statistical method that applies linear or radial wave matrices to images, capturing the heterogeneity and uniformity of fibrotic endometrium 34 . This may be attributed to subtle changes in the endometrial microstructure resulting from varying degrees of fibrosis infiltration, which lead to changes in voxel values that remain undetectable to the naked eye on MRI images. Radiomics analysis showed that LR-based merged model overperformed ET+T2 model for differentiating between SEF, MMEF and HC in the training set and test set. This indicated that radiomics models had an advantage over endometrial quantitative assessment by radiologists. It is worth noting that MMEF patients showed relatively inferior discriminative ability compared to HCs and SEF patients in both ET+T2 model and merged model. This might be due to the small sample size. Besides, most MMEF patients in this study suffered from moderate endometrial fibrosis, which reduced the ability for the discrimination between MMEF and SEF. The present work reveals that radiomics models based on T2WI and T2 mapping could significantly enhance the non-invasive staging of endometrial fibrosis and address the limitations of conventional methods for assessing endometrial fibrosis on MRI images. Multiclass classifications are more useful and closer to the radiologist’s reasoning in the daily clinical practice 31 , 37 . Due to the invasiveness of hysteroscopy, radiomics models can provide a new method to non-invasively and accurately evaluate the degree of endometrial fibrosis and help clinicians to perform individualized anti-fibrotic treatment. Besides, radiomics models can detect the microstructural changes in patients with endometrial fibrosis, which can theoretically detect the endometrial recovery earlier than macroscopic morphological change for patients with effective treatment. Hence, radiomics models also has the potential for the dynamic follow-ups in patients with endometrial fibrosis. Our study had some limitations. First, the sample size was relatively small and imbalanced. However, compared with other similar studies focusing on liver and kidney fibrosis 23 , 38 our study remained with a larger patient number. Further prospective studies with larger and more balanced sample sizes will be required to validate its reproducibility. Second, our study was done based on data from one center and used a single 3.0T MRI scanner. Multi-center and different MRI scanners may be needed to test the generalizability of our developed radiomics models. Third, our three-class classification model distinguishes HC, MMEF, and SEF but does not directly evaluate pairwise differences between groups. While clinically meaningful, this may reduce sensitivity in differentiating intermediate stages. Future work will focus on binary classification with larger, balanced datasets to improve staging accuracy.

In conclusion, radiomics models based on T2WI and T2 mapping showed better classification performance and potential as a non-invasive predictor for the staging of endometrial fibrosis, which may help clinicians improve therapeutic strategies in endometrial regeneration and reproductive performance.

Endometrial fibrosis, a prevalent gynecological disorder, is one of the major cause of uterine-related infertility and significantly affects both the physical and psychological health of women of childbearing age 1 . Endometrial fibrosis is characterized by the formation of intrauterine adhesions (IUA) following endometrial injury, resulting in excessive deposition of extracellular matrix in the endometrial basalis layer 2 . Various treatments have been developed to reduce the incidence of endometrial fibrosis and increase the rate of pregnancy, including hysteroscopic lysis of adhesions, intrauterine contraceptive devices, solid and semi-solid (gel) barriers, and hormonal treatments 3 . While such treatments tend to be more effective for patients with mild to moderate endometrial fibrosis (MMEF), the recurrence rate can rise to as high as 20-62.5% in patients with severe endometrial fibrosis (SEF) 4 , 5 . In recent years, it has been reported that stem cells therapy could effectively reverse endometrial fibrosis and promote endometrial regeneration in SEF patients 6 – 8 . Therefore, accurate staging of endometrial fibrosis is crucial for formulating an appropriate treatment plan 9 , 10 . Although hysteroscopy remains the gold standard for assessing endometrial fibrosis, it is limited by its invasive nature and is unsuitable for routine clinical follow-ups 11 . Ultrasound and conventional magnetic resonance imaging (MRI) can both reveal the morphological changes of endometrium. It has been reported that T2-weighted imaging (T2WI) can show thin endometrium or low signal fibrotic scar in the uterine cavity 12 . However, these conventional technique cannot quantify the microenvironmental changes in fibrotic endometrium 13 , 14 . T2 mapping is a quantitative functional MRI technique and can evaluate the accumulation of water and blood in tissue by means of the parameter of T2 relaxation time (T2 value), which can indirectly reflect the microenvironmental changes 15 , 16 . T2 mapping has been used to evaluate liver and myocardial fibrosis 17 , 18 . Our previous study also confirmed the feasibility of T2 mapping in the staging of endometrial fibrosis 19 . However, endometrial fibrotic scar has an inhomogeneous distribution in the uterine cavity 20 . Although T2WI and T2 mapping has advantage in the morphological and microenvironmental evaluation of endometrial fibrosis, neither of the techniques can detect the heterogeneity of distribution in fibrotic endometrium pixel by pixel 19 . Therefore, T2WI or T2 mapping alone may be insufficient for accurate assessment of endometrial fibrosis. Radiomics can provide large amounts of high-dimensional features from medical images and has the potential to uncover additional microstructural changes that might not be detected by visual inspection of images 21 . Radiomics has been proven effective in accurately detecting and differentiating fibrosis in various organs, including liver, kidney and heart 22 – 24 . To our knowledge, the use of MRI radiomics to stage endometrial fibrosis has not been investigated. We hypothesized that radiomic features derived from T2WI and T2 mapping may improve the staging of endometrial fibrosis. Therefore, this study aimed to develop radiomics models based on T2WI and T2 mapping for the non-invasive staging of endometrial fibrosis and compare the efficacy of the optimal radiomics models to conventional MRI evaluation models.

Below is the link to the electronic supplementary material. Supplementary Material 1. Supplementary Material 1.