A Dual-Sequence MRI-Based Radiomics Model for Predicting High-Intensity Focused Ultrasound Ablation Efficacy in Adenomyosis Treatment

Adenomyosis is a prevalent gynecological disorder affecting women of reproductive age, with an incidence rate ranging from 7% to 23%. 1 The disorder is characterized by the invasion of endometrial tissue into the uterine muscle layer, leading to uterine enlargement, dysmenorrhea, and menorrhagia. 2 Initial diagnosis is primarily based on clinical symptoms, with magnetic resonance imaging (MRI) being the most accurate radiological tool for diagnosing adenomyosis. 3–5 Adenomyosis treatment options include pharmacotherapy, surgical intervention, and interventional procedures. 6–11 Although conservative surgery has proven effective in over 50% of patients, long-term follow-up data remains insufficient. 12 , 13 High-intensity focused ultrasound (HIFU) is a novel non-invasive treatment modality that uses ultrasound’s penetrative, directional, and focal characteristics to induce coagulative necrosis within targeted areas, destroying lesions without damaging surrounding tissues. 14–16 HIFU treatment alleviates symptoms, with its efficacy closely related to immediate postoperative ablation rates. 17 Imaging biomarkers, such as T2-weighted MR imaging signal intensity, apparent diffusion coefficient, and ultrasound perfusion parameters, are commonly utilized to predict ablation efficacy but with limited accuracy. 18 , 19 In current clinical practice, the non-perfused volume ratio (NPVR) serves as an index for evaluating HIFU efficacy. 20 Advances in imaging technology, particularly the extraction of imaging features to construct accurate and objective efficacy prediction models, are crucial for the clinical expansion of HIFU, potentially benefiting more patients with adenomyosis. 21 The T2-Fat Suppression (T2FS) sequence provides information on the aqueous components within lesions, helping to identify areas of necrosis or edema, while the T1 contrast-enhanced (T1C) sequence reflects the blood supply and perfusion status of the lesions, aiding in the assessment of the non-perfused region post-treatment. 22 , 23 In evaluating HIFU ablation efficacy for adenomyosis, the combination of these two sequences offers a comprehensive reflection of the lesion’s physical characteristics and treatment response. Specifically, T2FS illustrates structural changes within the lesion, while T1C highlights the blood flow and perfusion status of the tissue after ablation. Radiomics analysis uses comprehensive methods to quantify texture, shape, and intensity patterns within the tumor region on MRI images. It has been utilized for predicting early recurrence (within 2 years) after curative resection of hepatocellular carcinoma (HCC), 23 , 24 and for predicting residual myoma regrowth within one year in uterine myomas treated with HIFU ablation. 25 Additionally, previous studies have evaluated the combined predictive value of radiomics based on different MRI sequences in determining the non-perfused volume ratio (NPVR) following HIFU ablation for uterine fibroids. 26–28 This study primarily investigates the value of radiomics models based on dual-sequence MRI in predicting the efficacy of HIFU ablation in patients with adenomyosis.

Clinical and radiological characteristics of all study patients were analyzed using ablation rate thresholds of 70% and 50% ( Tables 1 and S2 ). For predicting an effective ablation rate greater than 50%, univariate analysis identified VAS_BeforeTreatment, Ultrasonic Blood Flow, Uterine Size, and Lesion Size as significant factors. Multivariate analysis showed that VAS_BeforeTreatment and Ultrasonic Blood Flow were independent predictors, with p-values below 0.05. For predicting an ablation rate greater than 70%, Age, VAS_BeforeTreatment, Ultrasonic Blood Flow, and Lesion Size were significant in univariate analysis, while VAS_BeforeTreatment and Ultrasonic Blood Flow remained independent predictors in multivariate analysis. Table 1 Baseline Clinicoradiological Characteristics of Patients with Adenomyosis in the Training and Testing Cohorts (Ablation Rate Threshold = 70%) Characteristics Training Group Testing Group Negative (n=57) Positive (n=22) p Negative (n=25) Positive (n=10) p Diseases (%) Adenomyosis 54 (94.7) 19 (86.4) 0.432 24 (96.0) 8 (80.0) 0.190 Other 3 (5.3) 3 (13.6) 1 (4.0) 2 (20.0) Lesion Type (%) Focal lesions 20 (35.1) 6 (27.3) 0.508 6 (24.0) 3 (30.0) 0.694 Diffuse lesions 37 (64.9) 16 (72.7) 19 (76.0) 7 (70.0) Menstruation Before Treatment (%) Moderate 3 (5.3) 2 (9.1) 0.178 2 (8.0) 1 (10.0) 1.000 Light 34 (59.6) 13 (59.1) 11 (44.0) 5 (50.0) Heavy 20 (35.1) 7 (31.8) 12 (48.0) 4 (40.0) Menstruation After Treatment (%) Moderate 13 (22.8) 4 (18.2) 0.298 6 (24.0) 5 (55.6) 0.193 Light 36 (63.2) 16 (72.7) 18 (72.0) 4 (44.4) Heavy 8 (14.0) 2 (9.1) 1 (4.0) 0 (0.0) Ultrasonic Blood Flow (%) I 0 (0.0) 1 (6.2) 0.792 0 (0.0) 0 (0.0) 1.000 II 28 (68.3) 12 (75.1) 9 (56.2) 5 (62.5) III 12 (29.3) 2 (12.5) 6 (37.5) 3 (37.5) IV 1 (2.4) 1 (6.2) 1 (6.1) 0 (0.0) Age (mean±sd) 42.316 ± 4.231 41.455 ± 4.564 0.448 41.36 ± 5.964 41.8 ± 5.073 0.828 VAS Before Treatment (Median[Q1~Q3]) 8[6~8] 6[1.25~7.75] 0.017 6[1~8] 6[5~6.75] 0.941 VAS After Treatment (Median[Q1~Q3]) 3[1~3] 0.5[0~2] 0.010 0[0~2] 2[1.25~2.75] 0.047 Blood Platelets (Median[Q1~Q3]) 272[215~327] 249.5[205.25~296.5] 0.104 264[219~300] 237.5[222.25~259] 0.685 Hemoglobin (Median[Q1~Q3]) 119[99~131] 116.5[105.5~123.75] 0.581 114[100~121] 120[110.25~134.25] 0.369 PT (Median[Q1~Q3]) 11.4[10.8~12.2] 11.5[10.825~12] 0.943 11.3[10.8~12.3] 11.45[10.05~13.325] 0.826 PT. (Median[Q1~Q3]) 115.72[106.7~124.9] 113.35[104.15~115.6] 0.214 118.91[111.2~123.97] 108.68[99.055~125.85] 0.495 INR (Median[Q1~Q3]) 0.96[0.9~1.01] 0.95[0.903~0.97] 0.827 0.95[0.91~0.99] 0.95[0.835~1.02] 0.770 APTT (mean±sd) 27.146 ± 2.784 26.991 ± 2.912 0.832 27.668 ± 3.206 28.16± 1.930 0.235 F1B (mean±sd) 2.751 ± 0.438 2.565 ± 0.598 0.194 3.26 ± 2.584 2.792 ± 0.704 0.956 TT (Median[Q1~Q3]) 11.4[10.9~11.9] 11.4[11.025~12.275] 0.626 11.3[10.7~11.6] 11.3[10.65~12] 0.913 D.D (Median[Q1~Q3]) 0.07[0.07~0.08] 0.07[0.06~0.152] 0.995 0.07[0.07~0.16] 0.07[0.07~0.07] 0.335 CA125 (Median[Q1~Q3]) 71.44[40.68~109.95] 50.1[27.732~81.028] 0.159 51.19[31.095~85.635] 48.515[32.235~104.34] 0.515 Uterine Size (Median[Q1~Q3]) 189 303.775[132,717.253~222,012.38] 167 714.51[114,288.72~237 746.179] 0.528 255 656.64[176,780.94~309,804.07] 184 455.4[151,356.02~200,968.13] 0.013 Lesion Size (Median[Q1~Q3]) 68 451[40,986~117 174] 31 839[23 478~50,658] 0.002 93,655.5[66,838.5~135 222] 80,856[43 719.847~107 829] 0.220 Baseline Clinicoradiological Characteristics of Patients with Adenomyosis in the Training and Testing Cohorts (Ablation Rate Threshold = 70%) In this study, 2264 features were extracted from the T2-FS and T1C sequences. Specific image filters, radiomics features categories and quantity statistics can be found in Table S1 . For the intra-cohort correlation coefficient analysis, a threshold of 0.75 was set, resulting in 1827 reproducible radiomics features in T2-FS and 1618 in T1C. Following correlation and Lasso feature selection on the training set, 11 features were identified in both sequences when the ablation rate threshold was set at 70% (see 13). At a threshold of 50%, 14 features remained in T2-FS and 12 in T1C (see Table S3 ). Figure S1 shows the Lasso path diagrams for the two MR sequences at different ablation rates. In the Clinical model, the AUC achieved on the test set was 0.6, indicating poor performance (see Table 2 , Figure 3A and B ). The T1C model and T2FS model exhibited similar performance, with test set AUCs of 0.709 and 0.718, respectively. The Radiomics model, which combined T2FS and T1C features (see Table S4 ) at the feature level, achieved an AUC of 0.840. The Combination model, which incorporated clinical features along with T2FS and T1C, achieved an AUC of 0.738, performing slightly worse than the Radiomics model. However, the Radiomics model did not show a significant improvement over the single MR sequence models in both the training and test sets ( Table 3 ). Table 2 Performance of Radiomics Features and Fusion Models in Predicting the Achievement of 50% and 70% Ablation Rates in Adenomyosis Patients Ablation rates threshold Models AUC (95% CI) Accuracy Sensitivity Specificity Precision F1-score Train Test Train Test Train Test Train Test Train Test Train Test 70% Clinical 0.79(0.679–0.902) 0.6(0.399–0.801) 0.797 0.600 0.545 0.700 0.895 0.560 0.667 0.389 0.600 0.500 T1C 0.930(0.867–0.993) 0.770(0.585–0.955) 0.861 0.714 0.864 0.800 0.860 0.680 0.704 0.500 0.776 0.615 T2FS 0.954(0.914–0.994) 0.676(0.455–0.897) 0.899 0.800 0.864 0.500 0.912 0.920 0.792 0.714 0.826 0.588 Radiomics 0.986(0.967–1.000) 0.804(0.614–0.994) 0.899 0.857 0.682 0.600 0.982 0.960 0.938 0.857 0.790 0.706 Combination 0.994(0.985–1) 0.736(0.524–0.948) 0.911 0.800 0.682 0.600 0.999 0.880 0.999 0.667 0.811 0.632 50% Clinical 0.799(0.702–0.897) 0.6(0.399–0.802) 0.759 0.629 0.884 0.524 0.611 0.786 0.731 0.786 0.800 0.629 T1C 0.864(0.787–0.941) 0.709(0.528–0.891) 0.772 0.714 0.837 0.857 0.694 0.500 0.766 0.720 0.800 0.783 T2FS 0.881(0.808–0.953) 0.718(0.548–0.887) 0.823 0.686 0.977 0.524 0.639 0.929 0.764 0.917 0.857 0.667 Radiomics 0.889(0.819–0.958) 0.840(0.706–0.974) 0.810 0.771 0.721 0.762 0.917 0.786 0.912 0.842 0.805 0.800 Combination 0.926(0.868–0.983) 0.738(0.557–0.919) 0.873 0.771 0.953 0.81 0.778 0.714 0.837 0.810 0.891 0.810 Notes : Bold values are the performance of the radiomics model, which integrated T2FS and T1C features at the feature level demonstrated strong predictive performance, achieving AUC values of 0.804 (at 70% ablation rate threshold) and 0.840 (at 50% ablation rate threshold). Table 3 Comparison of Models in Training and Test Cohorts (Delong Test, NRI and IDI Test) Ablation rates threshold Compared models Basic model Delong test (p value) NRI | p value IDI | p value Training cohort Testing cohort Training cohort Testing cohort Training cohort Testing cohort 70% Radiomics Clinical 0.001 0.085 0.416 | 0.001 0.580 | 0.001 0.382 | < 0.001 0.260 | 0.005 T1C 0.048 0.682 0.164 | 0.071 0.080 | 0.652 −0.052 | 0.296 −0.024 | 0.772 T2FS 0.047 0.039 0.151 | 0.087 0.080 | 0.597 0.052 | 0.296 0.024 | 0.772 Combination 0.359 0.086 0.010 | 0.899 0.120 | 0.065 0.152 | 0.001 0.135 | 0.001 50% Clinical 0.128 0.067 0.079 | 0.524 0.405 | 0.064 0.093 | 0.180 0.154 | 0.254 T1C 0.541 0.173 0.101 | 0.349 0.191 | 0.273 −0.021 | 0.697 0.038 | 0.679 T2FS 0.877 0.309 0.032 | 0.814 0.238 | 0.307 −0.503 | 0.485 0.064 | 0.550 Combination 0.166 0.249 −0.066 | 0.532 0.405 | 0.011 0.083 | 0.017 0.167 | 0.006 Notes : Bold values are statistically significant with p0 and integrated discrimination improvement (IDI) >0 were positive improvement, indicating that the predictive ability of the new model was better than the old one. Figure 3 ROC curves for the task of predicting ablation rates greater than 50% and 70%. ( A , C ) training cohort, ( B , D ) testing cohort. Performance of Radiomics Features and Fusion Models in Predicting the Achievement of 50% and 70% Ablation Rates in Adenomyosis Patients Notes : Bold values are the performance of the radiomics model, which integrated T2FS and T1C features at the feature level demonstrated strong predictive performance, achieving AUC values of 0.804 (at 70% ablation rate threshold) and 0.840 (at 50% ablation rate threshold). Comparison of Models in Training and Test Cohorts (Delong Test, NRI and IDI Test) Notes : Bold values are statistically significant with p0 and integrated discrimination improvement (IDI) >0 were positive improvement, indicating that the predictive ability of the new model was better than the old one. ROC curves for the task of predicting ablation rates greater than 50% and 70%. ( A , C ) training cohort, ( B , D ) testing cohort. In the Clinical model, the AUC achieved on the test set was 0.6 (see Table 2 , Figure 3C and D ). T1C exhibited higher classification efficiency at the model level than T2FS (test AUC in RF: 0.77 vs 0.676, p=0.45). The Radiomics model, which combined T2FS and T1C features at the feature level, achieved an AUC of 0.804. The Combination model, which incorporated clinical features along with T2FS and T1C, achieved an AUC of 0.736, performing slightly worse than the Radiomics model. In both training and test sets, the Radiomics model significantly outperformed the T2-FS model (Delong test p-values were 0.047 and 0.039 as shown in Table 3 ). However, no significant difference was observed between the fusion and T1C models. ( Table 3 ). The calibration curves based on random forest prediction results displayed a better fit between the predictive model for whether the ablation rate exceeds 50% and the actual ablation rate in both the training set cohort and the test cohort compared to ablation rate surpasses 70% (refer to Figure S2 ). The p-values for the Hosmer-Lemeshow test on the training and test sets were 0.286, 0.372, 0.269, and 0.112, 0.148, and 0.594 for the T1C, T2FS, and Radiomics model, respectively, based on the Random Forest classifier (task 50%). The decision curves illustrate that the combination model yielded a greater net benefit than the single-sequence model across a wider range of risk thresholds (see Figure 4 ). Figure 4 Decision curves for the task of predicting ablation rates greater than 50%. ( A and B ) training cohort, ( C and D ) testing cohort. Decision curves for the task of predicting ablation rates greater than 50%. ( A and B ) training cohort, ( C and D ) testing cohort. Figure S1 exhibits the correlation analysis between the radiomics features and the prediction label regarding whether the ablation rate exceeds 50%. Table S5 and Figure S3 reveal that eight radiomics features showed significant correlations with the label and exhibited significant differences between the cohort with an ablation rate greater than 50% and the cohort with an ablation rate less than 50%.

The Institutional Review Board approved this retrospective analysis, and informed consent was not required (IRB no. 2024–056). Between July 2021 and July 2023, 188 individuals underwent HIFU ablation at our center. Eligibility for participation included: (1) women between the ages of 18 and the menopausal transition with diagnosed adenomyosis through clinical and radiological assessments; (2) individuals undergoing their initial HIFU ablation for adenomyosis; (3) MRI scans performed within three days prior to and following the ablation; (4) no previous surgical intervention or medical treatment for adenomyosis. Participants were excluded if they had: (1) other identified uterine or adnexal issues; (2) inadequate MRI image quality. A total of 114 patients who completed a one-year follow-up were included in this study for further analysis. Patients were randomly divided into a training set and a test set at a ratio of 7:3. The patient recruitment flowchart is presented in Figure 1 . Figure 1 Study flowchart of the enrolled patients with exclusion criteria. Study flowchart of the enrolled patients with exclusion criteria. Patients underwent examination using a 1.5T MRI scanner (MAGNETOM Aera, Siemens Healthcare, Germany) equipped with a body phased-array coil. The primary imaging sequences selected were axial T2-weighted MR imaging (T2WI) with fat saturation and contrast-enhanced T1-weighted MR imaging (CE-T1WI). The imaging parameters were set as follows: (1) For T2WI with fat saturation, the repetition time (TR) was set at 2500 ms, and the echo time (TE) was set at 88 ms. The slice thickness was 4 mm, the slice spacing was 1.2 mm, and the field of view (FOV) was 260 mm × 260 mm. (2) For CE-T1WI, the TR was set at 6.8 ms, and the TE was set at 2.4 ms. The slice thickness was 3 mm, the slice spacing was 0.6 mm, and the FOV was 260 mm × 260 mm. Dimeglumine gadopentetic acid was administered intravenously at a rate of 2.0 mL/s and a dose of 0.2 mmol/kg via a high-pressure syringe from the superficial vein of the forearm. The injection contained 20 mL of gadobenate dimeglumine, equivalent to 6.680g of gadobenic acid and 3.900g of N-methylglucamine, provided by Shanghai Bolaike Xinyi Pharmaceutical Co., Ltd. CE-T1WI observations were made 35 seconds after the injection. All original axial T2WI and CE-T1WI images before and after HIFU treatment were transferred to the offline image workstation. Prior to HIFU treatment, Radiologist 1 (with 5 years of experience in gynecologic tumor diagnosis) manually mapped each patient’s T2WI and CE-T1WI layer by layer, following the edge of the relatively low-signal lesions. This mapping referenced the multi-sequence MR enhancement images of the patient’s pelvis in the PACS system, with regions of interest (ROIs) encompassing the entire tumor. Following HIFU treatment, Radiologist 1 manually traced the lesion of each patient along the edge of the ablation area (necrotic area), with ROIs covering the entire ablation zone. These ROIs were then reviewed by Radiologist 2 (with 15 years of experience in gynecologic tumor diagnosis). All patients underwent MRI within 3 days after HIFU to assess the endometrium’s condition and evaluate the non-perfused volume (NPV). The volume of the adenomyosis lesion and the NPV were measured on each slice of the contrast-enhanced T1WI. The ablation volume (V) was calculated using the ellipsoid volume formula: V = (A × B × C) / 6, where A, B, and C represent the length, width, and height of the lesion, respectively. 29 The non-perfused volume ratio (NPVR) was determined as NPV/adenomyosis × 100%. Studies have shown that the efficacy of HIFU ablation for adenomyosis correlates with the NPVR. 30 Previous research has indicated that satisfactory clinical outcomes can be achieved when the NPVR is 50% or higher. Multiple studies have reported that the average or median NPVR for HIFU ablation of adenomyosis ranges from 50% to 70%. 30–33 Therefore, in this study, to identify an optimal model with NPVR thresholds of 50% and 70%, adenomyosis cases in the training and test cohorts were categorized into effective ablation (NPVR ≥ 70% and ≥ 50%) and ineffective ablation (NPVR < 70% and < 50%) cohorts. Clinical and radiological features that might affect the non-perfused volume ratio (NPVR) of adenomyosis were analyzed. These included age, preoperative blood routine, coagulation function, Carbohydrate Antigen 125 (CA-125), size and type of adenomyosis lesions, and Color Doppler flow imaging (CDFI) signals. The Visual Analog Scale (VAS) scores and menstrual flow were assessed before and after treatment. The CDFI signal was categorized using the Adler grading system as follows: Grade 0 indicates no short rod or punctate flow signals; Grade I indicates one or two short rods or punctate flow signals; Grade II indicates three to four punctate vessels; Grade III indicates more than four blood vessels or an intertwined network of blood vessels. The VAS score was used to evaluate changes in dysmenorrhea. Menstrual pain was scored on a 0–10 Numerical Rating Scale (NRS), where 0 indicates no pain, 1–3 points indicate mild pain, 4–6 points indicate moderate pain, and 7–10 points indicate severe pain. Radiomics analysis, 34 including tumor segmentation, was performed using the uAI Research Portal (uRP, version 231115, United Imaging Intelligence Co., Ltd., China). The workflow for radiomics analysis consists of the following steps: tumor segmentation, feature extraction, feature selection, model construction, model analysis, and evaluation. A flowchart depicting this process is provided in Figure 2 . Figure 2 Study flowchart of Radiomics analysis. Study flowchart of Radiomics analysis. For each subject, an abdominal radiologist with 10 years of experience manually delineated the tumor region (volume of interest, VOI) based on dual MR sequences T2FS and T1C). A senior abdominal radiologist reviewed these delineations. To evaluate the reproducibility of radiomics features, 30 patients with adenomyosis were randomly selected and outlined by Radiologist 2. The intra-class correlation coefficient (ICC) was utilized to assess the consistency between the delineations of the two cohorts. To reduce heterogeneity among MR images, all MR images were resampled using bilinear interpolation to achieve an isotropic voxel size of 1 × 1×1 mm 3 . Subsequently, intensity normalization was performed by applying a fixed bin width of 25 and z-score normalization to achieve a standard normal distribution of image intensities. Radiomics features were extracted using the uAI Research Portal, which integrates PyRadiomics ( https://pyradiomics.readthedocs.io/en/v3.0.1/ ). For each sequence, 2264 radiomics features were extracted and categorized into first-order, shape, texture, and higher-order features ( Table S1 ). Feature selection was conducted exclusively on the training set to prevent information leakage between the training and test datasets. Radiomics features were first extracted and subsequently normalized using z-score normalization. Following this, features with an intra-class correlation coefficient (ICC) of 0.75 or higher in test-retest evaluations were identified as reproducible and selected for further analysis. Optimal predictive features were then identified through correlation analysis (with a p-value ≤ 0.05) and the least absolute shrinkage and selection operator (LASSO) method. In both classification tasks, the optimal feature selection results were used to construct five models: clinical model, T1C model (single-sequence radiomics model), T2FS model (single-sequence radiomics model), Radiomics model (combination of T1C and T2FS), and the Combination model (combination of T1C, T2FS, and clinical features). For all fusion models, we employed feature-level fusion, merging the features from the single-sequence models and selecting them for the fusion model. The model’s performance was evaluated by generating receiver operating characteristic (ROC) curves and calculating the area under the curve (AUC), as well as metrics for sensitivity, specificity, accuracy, precision, and F1 scores. The Delong test with bootstrap resampling was used to compare predictive efficiency among the models, and false discovery rates (FDR) were adjusted using the Benjamini-Hochberg method. Additionally, the performance of various Radiomics models was assessed using the net reclassification improvement index (NRI) and integrated discrimination improvement index (IDI). The actual and predicted ablation rates at thresholds of 50% and 70% were evaluated using the Hosmer-Lemeshow test, and calibration curves were plotted. To validate the clinical applicability of Radiomics models, decision curves were constructed to quantify the net benefits under various risk thresholds. Student’s t-tests were conducted for variables with a normal distribution, while the Mann–Whitney U -test was used for variables with a non-normal distribution. The chi-square test was applied to qualitative variables to determine statistically significant differences. Univariate and multivariate analyses were employed to identify independent predictors. The statistical analysis was carried out using R software, version 4.1.3. All statistical tests were two-sided, with a p-value of less than 0.05 considered statistically significant.

This study is the first to predict the efficacy of HIFU ablation for adenomyosis based on T1C combined with T2FS radiomics, which will help clinicians judge the difficulty of ablation before surgery, select an appropriate treatment plan for patients, and promote the development of personalized medicine. The prediction model showed better alignment with actual ablation rates when predicting rates exceeding 50%. Combining radiomics features from two MR sequences achieved an AUC of 0.84 in the testing set. The decision curves illustrate that the combination model yielded a greater net benefit than the single-sequence radiomics model across a wider range of risk thresholds. In evaluating HIFU ablation efficacy for adenomyosis, the combination of these sequences provides a comprehensive view of the lesion’s structure and treatment response, with T2FS highlighting structural changes and T1C showing post-ablation perfusion status. We included T1C sequences in this study to improve the characterization of adenomyosis by the model, which might also be one of the reasons for the high efficiency of the Radiomics model. Previous studies have also confirmed that the combination of sequences could fully reflect the information of tumors. 35 Recent studies have evaluated the value of radiomics models based on non-contrast-enhanced MRI in predicting HIFU ablation efficacy for uterine fibroids. 36 , 37 They also introduced the potential of dual-sequence MRI radiomics combination models for predicting ablation rates. These combination models demonstrate superior predictive performance compared to radiomics and clinical-radiological models. Similarly, in this study, a clinical model was also developed, but it did not show comparable classification performance to the Radiomics model in either the 50% or 70% efficacy prediction tasks. For the 50% ablation task, the Radiomics model performed better, with volume and texture features (especially gray-level co-occurrence matrix features) showing the strongest correlation with ablation rates and effectively distinguishing treatment outcomes. T2FS signal intensity and morphological features also played a significant role in differentiating effective from ineffective ablation, with signal intensity particularly useful in predicting edema and necrosis. T1C contrast and uniformity were key in identifying areas of effective ablation, potentially helping assess treatment success. The combination of these features provides a more accurate prediction of ablation efficacy, particularly in HIFU treatment for adenomyosis (see Figures S3 and S4 ). This study has several limitations worth noting. Firstly, it adopts a single-center retrospective design with a small sample size. Validation through analysis of larger external datasets is required before implementation in clinical settings. Future studies include creating a radiomics model incorporating multiple imaging sequences and clinical information from various centers.

The combination model based on dual-sequence MRI radiomics can be used to predict the efficacy of HIFU treatment for adenomyosis. The results indicate that the combination model yields greater net benefit over single-sequence radiomics models. It was also found that the prediction model for ablation rates greater than 50% will help clinicians better select patients who benefit most from HIFU treatment, provide a reference for treatment decisions, and formulate accurate treatment plans.