{"paper_id":"b00fcd59-4268-4c1b-b0d5-2ead1335f8da","body_text":"International Journal of Hyperthermia\nISSN: 0265-6736 (Print) 1464-5157 (Online) Journal homepage: www.tandfonline.com/journals/ihyt20\nPrediction of clinical outcome for high-intensity\nfocused ultrasound ablation of adenomyosis\nbased on non-enhanced MRI radiomics\nZiyi Liu, Ziyan Liu, Xiyao Wan, Yuan Wang & Xiaohua Huang\nTo cite this article: Ziyi Liu, Ziyan Liu, Xiyao Wan, Yuan Wang & Xiaohua Huang (2025)\nPrediction of clinical outcome for high-intensity focused ultrasound ablation of adenomyosis\nbased on non-enhanced MRI radiomics, International Journal of Hyperthermia, 42:1, 2468766,\nDOI: 10.1080/02656736.2025.2468766\nTo link to this article:  https://doi.org/10.1080/02656736.2025.2468766\n© 2025 The Author(s). Published with\nlicense by Taylor & Francis Group, LLC\nPublished online: 23 Feb 2025.\nSubmit your article to this journal \nArticle views: 2124\nView related articles \nView Crossmark data\nCiting articles: 3 View citing articles \nFull Terms & Conditions of access and use can be found at\nhttps://www.tandfonline.com/action/journalInformation?journalCode=ihyt20\n\nInternatIonal Journal of HypertHermIa\n2025, Vol. 42, no . 1, 2468766\nPrediction of clinical outcome for high-intensity focused ultrasound ablation \nof adenomyosis based on non-enhanced MRI radiomics\nZiyi Liu *, Ziyan Liu *, Xiyao Wan, Yuan Wang and Xiaohua Huang \nDepartment of r adiology, a ffiliated Hospital of north Sichuan m edical College, nanchong, China\nABSTRACT\nObjectives:  The study aimed to develop a non-enhanced MRI-based radiomics model for the \npreoperative prediction of the efficacy of adenomyosis after high-intensity focused ultrasound (HIFU) \ntreatment.\nMethods: The data of 130 patients with adenomyosis who underwent HIFU treatment were reviewed. \nBased on a non-perfused volume ratio (NPVR) of 50%, the patients were assigned to high ablation rate \nand low ablation rate groups. A radiomics model was constructed from the screened radiomics features \nand its output probability was calculated as the radiomics score (Radscore). The clinical-imaging model \nwas constructed from the independent predictors of clinical-imaging characteristics. The combined \nmodel was constructed by integrating Radscore and clinical-imaging independent predictors. Receiver \noperating characteristic (ROC) curves, the Delong test, and decision curve analysis (DCA) were used to \nevaluate the models.\nResults: The combined model had the best overall performance among the three models. The AUC \n(95% CI), specificity, sensitivity, accuracy, and precision of the combined model were 0.860 (0.786–0.935), \n0.780, 0.756, 0.769, 0.738 in the training set, and 0.878 (0.774–0.983), 0.859, 0.667, 0.769, 0.800 in the \ntest set, respectively. The Delong test showed that the performance of both the radiomics and combined \nmodels differed significantly from the clinical-imaging model. But the performance of the combined and \nthe radiomics model was statistically equivalent. The DCA indicated that the combined model had \nbetter clinical net benefit.\nConclusion: The combined model based on non-enhanced MRI radiomics was effective in predicting \nthe outcome of HIFU ablation of adenomyosis before surgery.\n1.  Introduction\nAdenomyosis is a common benign gynecological disease in \nwhich the endometrial glands and mesenchyme invade and \ngrow into the myometrium of the uterus [ 1,2]. It is primarily \ncharacterized by the following symptoms: dysmenorrhea, \nabnormal menstrual cycles, increased vaginal discharge, infer -\ntility, and abnormal menstrual flow, which significantly affect \nthe patient’s quality of life [ 3,4]. Traditional treatments for \nadenomyosis, such as medication or surgery, have limited \neffectiveness due to the disease’s chronic estrogen-dependent \nnature and as the lesions tend to be poorly differentiated \nfrom the normal myometrium or are diffusely distributed, \nresulting in a high recurrence rate [ 3,5]. Hysterectomy is a \nradical treatment; however, it is not the best choice for \nwomen who want to keep their uterus or are trying to con -\nceive [ 6]. Therefore, high-intensity focused ultrasound (HIFU), \na noninvasive ablation procedure, is often used to treat ade -\nnomyosis [ 7–9]. Its benefits include uterine preservation, \nsafety, efficacy, and minimal side effects [ 7,10]. However, due \nto individual differences and variations in the lesion tissue, \nnot all adenomyosis treatments are equally effective. Currently, \nthe most popular way to measure the short-term effective -\nness of HIFU is the non-perfused volume ratio (NPVR) [ 11,12]. \nThe NPVR is associated with volume reduction and symptom \nrelief after treatment [ 13–15]. The accurate prediction of the \nNPVR after HIFU treatment is therefore important for the \nselection of suitable patients, cost-saving, and the develop -\nment of treatment plans.\nMRI is a significant tool for the diagnosis and evaluation of \nadenomyosis [16,17]. Several studies have found that imaging \nfeatures such as T2WI signal intensity, the number of T2 \nhigh-signal lesions, and the kind of T1WI enhancement can be \nused to predict the NPVR when using HIFU [ 6,18,19]. However, \ntraditional clinical-imaging features have shown limited predic-\ntive power (AUC = 0.720) [ 19]. Moreover, the visual interpreta -\ntion of medical images is dependent on the experience of the \nobserver and the results thus lack objectivity. Radiomics can \nextract quantitative image features that are not apparent to \nhuman eyes to establish relevant predictive models [ 20,21]. \nMost previous radiomics-based studies [2,22] on the prediction \nof HIFU efficacy in treating adenomyosis have focused on single \n© 2025 t he a uthor(s). p ublished with license by taylor & f rancis Group, ll C\nCONTACT Xiaohua Huang  15082797553@163.com   Department of r adiology, a ffiliated Hospital of north Sichuan m edical College, no. 1, m aoyuan South \nr oad, Shunqing District, nanchong 637000, China.\n*t hese authors have contributed equally to this work.\nhttps://doi.org/10.1080/02656736.2025.2468766\nt his is an o pen a ccess article distributed under the terms of the Creative Commons a ttribution license ( http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, \ndistribution, and reproduction in any medium, provided the original work is properly cited. t he terms on which this article has been published allow the posting of the a ccepted \nmanuscript in a repository by the author(s) or with their consent.\nARTICLE HISTORY\nr eceived 15 o ctober 2024\nr evised 27 January 2025\na ccepted 13 f ebruary 2025\nKEYWORDS\nmagnetic resonance \nimaging; radiomics; \nadenomyosis; high-intensity \nfocused ultrasound; \nprediction\n\n2 Z. LIU ET AL.\nT2WI images, and did not consider the differences and comple -\nmentarity between different sequences. Although CE-T1WI \nimages can provide blood perfusion data, the injection of gad -\nolinium contrast agents not only increases the financial burden \non the patient, but may also have some adverse effects. \nDWI-based ADC maps can provide an effective reflection of the \ndegree of diffusion of tissue water molecules [23], which can be \nused as an effective complement to T2WI sequences.\nTherefore, this study aimed to construct a radiomics model \nbased on non-enhanced images (ADC and T2WI images) to pre-\ndict the NPVR in the HIFU treatment of adenomyosis. \nFurthermore, this study combined radiomics and clinical-imaging \nfeatures to improve the predictive performance.\n2.  Materials and methods\n2.1.  Patients\nRetrospective data were collected from 336 patients who under-\nwent HIFU for adenomyosis at the Affiliated Hospital of North \nSichuan Medical College between September 2021 and \nDecember 2023. The inclusion criteria were: (1) women experi -\nencing symptoms of adenomyosis; (2) no prior history of surgery \nor medication therapy related to the condition; (3) an MRI scan \nconducted no more than three days before and after ablation \ntreatment; (4) a diagnosis of adenomyosis confirmed by clinical \nand radiological examinations. The exclusion criteria were: (1) \ninsufficient or absent imaging data; (2) the presence of other \ngynecological disorders, such as pelvic inflammatory disease or \nuterine fibroids; (3) poor image quality affecting the drawing of \nthe target region; (4) pregnancy or breastfeeding; (5) below the \nage of 18 years. Figure 1 shows that 130 patients who met the \ninclusion and exclusion criteria for the research were included in \nthe cohort. The patients’ ages ranged from 45 to 48 years.\n2.2.  MRI scanning protocol\nThe uMR790 3.0 T, 12-channel body phased array coil from \nUnited Imaging was used to conduct the pelvic MRI scans. To \nminimize artifacts caused by respiratory motion, the patient \nwas given breathing training before the test and a bandage \nwas used to compress the abdomen. GD-DTPA was injected \nat a flow rate of 1.0 ml/s, with a dosage of 0.1 mmol/kg used \nto achieve enhanced scanning. The arteries were imaged at \n15, 30, and 45 s after injection of the GD-DTPA, representing \nthe early, middle, and late time points, respectively. The \nb-values, diffusion sensitivity coefficients, were set at 50 and \n800 s/mm2. Table 1 provides details of the scanning order and \nprimary parameters.\n2.3.  Patient grouping\nWhen analyzing adenomyosis lesions that were found to be \ngenerally regular, the sagittal CE-T1WI and T2WI images were \nused to assess the postoperative ablation volume and adeno -\nmyosis volume, respectively. The ellipsoid formula (0.5233 × \nlongitudinal diameter × anteroposterior diameter × transverse \ndiameter) [ 24] was used to calculate the non-perfused vol -\nume (NPV) and adenomyosis volume (V) of the lesion. For \ndiffuse adenomyosis or lesions with unclear boundaries, man -\nual layer-by-layer delineation of the region of interest (ROI) \nwas performed, with automatic determination of V and NPV \nusing the 3D-slicer software.\nWe estimated the ablation rate as NPVR, which is equal to \nNPV/V*100%. Using NPVR values of 50% as the threshold \n[25], the patients were allocated to high ablation rate (NPVR \n≥ 50%, n = 59) and low ablation rate (NPVR< 50%, n = 71) \ngroups. The groups were then randomly divided into training \nand test sets with a ratio of 7:3. The training set included 50 \ncases with low ablation and 41 cases with high ablation, \nwhile the test set comprised 21 cases with low ablation and \n18 cases with high ablation.\n2.4.  Clinical–imaging features\nClinical-imaging features that were likely to affect the NPVR \nin adenomyosis were assessed. These included age, adeno -\nmyosis volume, type of adenomyosis (diffuse/focus), location \nof adenomyosis (anterior, posterior, anterior and posterior), \nlocation of uterus (anteverted, retroverted), the distance from \nFigure 1.  f low chart of patient recruitment.\n\nINTERNATIONAL  JOURNAL  OF HYPERTHERMIA 3\nthe anterior side of adenomyosis to skin, abdominal wall \nthickness, T2 signal intensity (hypointensity: lesions with \nlower signal intensity than the normal myometrium, isointen -\nsity: lesions with signal intensity similar to that of the normal \nmyometrium), the number of hyperintense foci on T2WI (few \nor multiple hyperintense foci: the number of hyperintense \npoints was ≤ or >5 on a single slice, respectively). Laboratory \ntest data were collected, including leukocyte count, red \nblood cell count, hemoglobin, and platelet count. HIFU treat -\nment parameters were collected, including treatment power, \ntreatment time and energy.\nUnivariate logistic regression was used to analyze the \nclinical-imaging characteristics to identify significant factors \n(p < 0.05). Multivariate logistic regression analysis was then \nperformed to identify the independent predictors associated \nwith NPVR after HIFU treatment.\n2.5.  Radiomics feature extraction and ROI segmentation\nTwo radiologists with 10 years’ experience in the diagnosis of \ngynecological disease and who were blinded to the specifics \nof the cases manually segmented the ROIs of the two \nsequences (ADC and T2WI), layer by layer using a 3D Slicer \n(version 5.6.1), making sure to include all of the adenomyosis \nlayers. The 3D volume of interest (VOI) of adenomyosis was \ngenerated by fuzing the ROIs of the different image layers. \nFor diffuse adenomyosis involving the entire uterine wall, the \nedge of the ROI was maintained at a specific distance from \nthe endometrial and plasma layers to avoid the involvement \nof the normal tissue. The original images were processed by \nLaplacian Gaussian filtering and wavelet transform filtering, \nand the radiomics features of each VOI were extracted using \nthe built-in radiomics plugin.\nTo guarantee the reproducibility of the radiomics features, \none-third of the ADC and T2WI images and outlined lesions \nwere re-analyzed. The consistency between the observers for \nthe extracted features was then assessed by computing the \nintergroup correlation coefficient (ICC). Only characteristics \nthat demonstrated strong consistency and stability (ICC > \n0.75) were retained, while the rest were discarded.\n2.6.  Radiomics feature selection\nThe following three steps were used for feature selection \nusing R (version 4.3.3) and uAI Research Portal (version 730): \n(1) The feature data were preprocessed using Z ⁃score normal -\nization to eliminate the dimensional effects of different fea -\ntures, (2) Features with variances below 0.8 were discarded \nusing variance thresholding, after which select K Best was \nused to remove features that did not show substantial differ -\nences between the two groups; (3) Least absolute shrinkage \nand selection operator (LASSO) was used to select the most \nrelevant features for analysis, which further reduced the \ndimensionality.\n2.7.  Model establishment\nThree models were constructed using logistic regression to \npredict the NPVR after HIFU treatment. The radiomics model \nwas constructed from the screened radiomics features and its \noutput probability was calculated as the radiomics score \n(Radscore). The clinical-imaging model was constructed from \nclinical-imaging features that had been found to be inde -\npendently predictive. The combined model was developed \nby integration of the Radscore and the independently predic -\ntive clinical-imaging features.\n2.8.  Evaluation of model performance\nThe receiver operating characteristic (ROC) curve and the \narea under the curve (AUC), specificity, sensitivity, accuracy, \nand precision, were used to evaluate the predictive efficacy \nof the different models. Differences between the models \nwere compared using the Delong test. The clinical benefits of \neach model were evaluated using decision curve analysis \n(DCA). To assess how well the model matched the data, cali -\nbration curves and the Hosmer-Lemeshow test were used. \nFigure 2  illustrates the workflow of the radiomics analysis.\n2.9.  Statistical analysis\nA statistical significance level of p < 0.05 was employed for \nthe analysis, which was conducted using SPSS (version 27.0) \nand R (version 4.3.3) software. The distribution of the quanti -\ntative data was determined using the Shapiro-Wilk test. Data \nthat followed a normal distribution are presented as ( xs± ), \nwhereas data that followed a skewed distribution are given \nas M (Q25, Q75). The independent samples t-test was used \nfor comparing normally distributed data, while the Mann-  \nWhitney U test was used for analyzing non-normally distrib -\nuted data. Qualitative data were analyzed using either Fisher’s \nexact test or the chi-square test.\n3.  Results\n3.1.  Clinical–imaging features\nThe training and test sets were compared based on their \nclinical-imaging characteristics ( Table 2 ). Hemoglobin exhib -\nited a statistically significant difference ( p = 0.041), whereas \nother characteristics did not ( p > 0.05).\nUnivariate logistic regression analysis indicated that loca -\ntion of adenomyosis, the number of hyperintense foci on \nT2WI and the distance from the anterior side of adenomyosis \nto skin differed significantly between the high ablation rate \ngroup and the low ablation rate group. Following this, multi -\nvariate logistic regression analysis determined the number of \nhyperintense foci on T2WI (OR = 0.337, p = 0.005) and  \nthe  distance from the anterior side of adenomyosis to skin \n(OR = 0.981, p = 0.037) were independent predictors of NPVR \n(Table 3).\nTable 1. m agnetic resonance sequences and parameters.\nparameters t2WI-fS Ce-t1WI DWI\ntr (ms) 3300 4.23 2863\nte (ms) 88.40 1.72 80\nt hickness (mm) 4 5 4\nSpacing (mm) 2 0 2\nfoV (mm) 260 × 260 350 × 300 280 × 260\nmatrix 256 × 256 304 × 75 128 × 100\nnote: tr: r epetition time; te: e cho time; foV: f ield of view.\n\n4 Z. LIU ET AL.\n3.2.  Radiomics features\nOf the overall 2446 radiomics features extracted from the \nADC and T2WI images, 1909 features (752 from ADC and \n1157 from T2WI) were retained according to the results of \nthe ICC test (ICC > 0.75). Ultimately, 11 radiomics features, \nincluding 5 from ADC images and 6 from T2WI images, were \nretained after the three screening steps.\nFigure 2.  f lowchart of radiomics.\nTable 2.  Comparison of clinical and imaging features between training and \ntest sets.\ntraining Set test Set p value\na ge (years) 43(39, 47) 43(36, 46) 0.326\nVolume (cm 3) 64.97(38.47, 112.27) 62.80(36.00, 108.35) 0.657\na bdominal wall \nthickness (mm)\n29.48 ± 7.65 27.88 ± 7.31 0.271\nDistance from the \nanterior side of \nadenomyosis to skin \n(mm)\n56.00(42.61, 78.77) 67.89(49.55, 73.89) 0.435\nl eukocyte (10 9·l −1 ) 5.97 ± 1.61 5.64 ± 1.35 0.263\nr ed blood (10 12·l −1 ) 4.36(4.02, 4.72) 4.27(4.01, 4.58) 0.162\nHemoglobin (g·l −1 ) 115(95, 130) 103(85, 124) 0.041*\nplatelet (10 9·l −1 ) 262(210, 320) 252(192, 327) 0.419\ntype of adenomyosis (n) 0.567\n Diffuse 28 14\n f ocus 63 25\nl ocation of the uterus (n) 0.292\n a nteverted 77 30\n r etroverted 14 9\nl ocation of adenomyosis \n (n)\n0.890\n a nterior 25 11\n p osterior 48 19\n a nterior and p osterior 18 9\nt2 signal intensity (n) 0.661\n Hypointensity 57 26\n Isointensity 34 13\nt he number of \nhyperintense foci on \nt2WI (n)\n0.788\n ≤5 42 17\n >5 49 22\ntreatment power (W) 400(400,400) 400(400,400) 0.933\ntreatment time (s) 614(374,905) 627(405,828) 0.847\nenergy (kJ) 361.20(245.60,466.80) 331.20(250.80,399.20) 0.863\n*p value represents the according parameter is of statistical significance.\nTable 3. univariate and multivariate logistic regression analyses between high \nand low ablation rate groups.\nunivariate multivariate\nor (95%CI) p or (95%CI) p\nage 1.027  \n(0.967, 1.090)\n0.386 —— ——\nVolume 1.000  \n(1.000, 1.000)\n0.112 —— ——\na bdominal wall \nthickness\n0.959  \n(0.915, 1.006)\n0.087 —— ——\nDistance from \nthe anterior \nside of \nadenomyosis \nto skin\n0.980  \n(0.963, 0.997)\n0.018* 0.981  \n(0.963, 0.999)\n0.037*\nl eukocyte 0.841  \n(0.666, 1.061)\n0.144 —— ——\nr ed blood 0.601  \n(0.297, 1.213)\n0.155 —— ——\nHemoglobin 1.001  \n(0.997, 1.005)\n0.608 —— ——\nplatelet 0.995  \n(0.979, 1.011)\n0.510 —— ——\nt ype of \nadenomyosis\n0.860  \n(0.410, 1.804)\n0.689 —— ——\nl ocation of \nuterus\n1.708  \n(0.668, 4.363)\n0.264 —— ——\nl ocation of \nadenomyosis\n0.350  \n(0.139, 0.883)\n0.026* 2.285  \n(0.756, 6.904)\n0.143\nt2 signal \nintensity\n1.573  \n(0.759, 3.259)\n0.223 —— ——\nt he number of \nhyperintense \nfoci on t2WI\n2.502  \n(1.231, 5.089)\n0.011* 0.337  \n(0.156, 0.726)\n0.005*\ntreatment power 1.017 (0.955, 1.039) 0.126 —— ——\ntreatment time 1.000 (0.999, 1.001) 0.889 —— ——\nenergy 1.000 (0.998, 1.003) 0.813 —— ——\n*p value represents the according parameter is of statistical significance.\n\nINTERNATIONAL  JOURNAL  OF HYPERTHERMIA 5\n3.3.  Evaluation of model performance\nThe results of the ROC curve analysis of the three models \nare shown in Figure 3  and Table 4 . The AUC of clinical-  \nimaging model, radiomics model, and combined model \nwere 0.692, 0.838, and 0.860 in the training set, and 0.646, \n0.868, and 0.878 in the test set, respectively. The  combined \nmodel showed the best overall performance among the \nthree models. Its AUC (95% CI), specificity, sensitivity, \n accuracy, and precision were 0.860 (0.786–0.935), 0.780, \n0.756, 0.769, and 0.738 in the training set, and  \n0.878 (0.774–0.983), 0.859, 0.667, 0.769, and 0.800 in the \ntest set.\nThe results of the Delong test showed that the perfor -\nmance of clinical-imaging model was significantly lower than \nthat of radiomics model ((training set: p = 0.048, test set: \np = 0.047)) and combined model (training set: p = 0.004, test \nset: p = 0.001). There was no significant difference between \nthe radiomics model and the combined model (training set: \np = 0.253, test set: p = 0.760).\nAccording to the DCA ( Figure 4A ), for the majority of the \nthreshold probabilities, the combined model provided greater \nclinical net benefit in predicting the NPVR of HIFU treatment. \nThe results of the Hosmer-Lemeshow test and the calibration \ncurves demonstrated that the combined model was \nwell-corrected (0.152 for the training set, 0.147 for the test \nset), as shown in Figure 4B .\n4.  Discussion\nIn this study, we constructed the clinical-imaging model, \nradiomics model and combined model, and we found that \nthe combined model that integrated the Radscore from the \nnon-enhanced MRI model along with the independent pre -\ndictors used in the clinical-imaging model had better predic -\ntive efficacy and clinical value. The AUC (95% CI), specificity, \nsensitivity, accuracy, and precision of the combined model \nwere 0.860 (0.786–0.935), 0.780, 0.756, 0.769, 0.738 in the \ntraining set, and 0.878 (0.774–0.983), 0.859, 0.667, 0.769, \n0.800 in the test set, respectively. This method will help clini -\ncians to predict the NPVR of HIFU treatment before surgery \nand screen patients suitable for HIFU treatment.\nRadiomics provides large amounts of information from \nmedical images and can identify heterogeneity in the spatial \ndistribution of lesions [ 26–28]. Radiomics-based MRI has  \nbeen used in studies on adenomyosis to predict the long  \nand short-term efficacy of HIFU and identify adenomyosis \n[1,2,22,29]. Li et  al. [ 2] utilized T2WI-based radiomics to predict \nthe long-term outcome of HIFU treatment in adenomyosis. The \nAUC, specificity, sensitivity, and accuracy of their \nradiomics-clinical model in the test set were 0.81, 0.71, 0.86, \nand 0.76, respectively. However, the sample size of their study \nwas small, with only 69 cases, and the model only incorpo -\nrated 4 radiomics features, which may have increased the risk \nof model overfitting and instability. Ying et  al. [ 22] constructed \nFigure 3.  roC curves for clinical-imaging model, radiomics model, and combined model in ( a ) the training and (B) test sets.\nTable 4. p erformance comparison of three models in training and test sets.\nauC (95% CI) Specificity Sensitivity a ccuracy precision\nClinical-imaging model training 0.692(0.583–0.802) 0.820 0.439 0.648 0.667\ntest 0.646(0.468–0.823) 0.857 0.278 0.590 0.625\nr adiomics model training 0.838(0.751–0.924) 0.840 0.683 0.769 0.778\ntest 0.868(0.756–0.980) 0.762 0.722 0.744 0.722\nCombined model training 0.860(0.786–0.935) 0.780 0.756 0.769 0.738\ntest 0.878(0.774–0.983) 0.859 0.667 0.769 0.800\n\n6 Z. LIU ET AL.\na model based on T2WI radiomics and deep learning to pre -\ndict adenomyosis lesion ablation by HIFU treatment. The AUC, \naccuracy, precision, recall, and F-score in the test set of their \nmodel were 0.861, 0.814, 0.832, 0.795, and 0.813, respectively. \nHowever, these studies focused only on radiomics features \nfrom a single T2WI image and did not address the variability \nand complementarity between different sequences.\nADC images obtained by the post-processing of DWI \nimages with different b-values can provide an accurate reflec -\ntion of the true diffusion properties of tissues by eliminating \nthe influence of the T2 transmission effect, presenting infor -\nmation on cells and the microcirculation [ 28,30,31]. Therefore, \nradiomics features that incorporate both T2WI and ADC \nsequences can more comprehensively and accurately reflect \nthe pathophysiology of adenomyosis, overcoming the limita -\ntion of the single sequence that provides only limited data \non the lesion. In this study, 11 features (6 from T2WI images \nand 5 from ADC images) were used to construct the radiom -\nics model. The results of this study suggest that both T2WI \nand ADC images are indispensable in predicting the NPVR for \nadenomyosis treatment with HIFU.\nThis study also analyzed the relationship between clinical-  \nimaging characteristics and NPVR. The number of hyperin -\ntense foci on T2WI and the distance from the anterior side of \nadenomyosis to skin were found to be independent predic -\ntors of NPVR in adenomyosis. A greater distance from the \nventral side of the adenomyosis to the skin and the presence \nof multiple hyperintense foci were not conducive to effective \nHIFU ablation in patients with adenomyosis, which is consis -\ntent with the findings of previous research [ 6,32]. By combin -\ning the independent clinical-imaging predictors and Radscore, \nincreased the overall performance and clinical benefit of the \ncombined model relative to those of the single radiomics \nand single clinical-imaging models. Consistent with the \nresults of recent studies, the predictive performance of the \nmodel improved when radiomics was combined with \nclinical-imaging features [ 2,33].\nThis study has several limitations. Firstly, due to the limited \navailability of data, external validation was not  performed. \nFurther research is necessary to acquire external validation data \nto confirm the model’s viability. Secondly, the study was \nretrospective with a small sample size and is thus potentially \nsubject to bias; further studies with larger sample sizes are \nneeded for verification of the findings. Finally, the ADC image \nmight not be as high-quality as the T2WI image. Even though \nthe outlining process used combined T2WI images to deter -\nmine the ROI and the ICC test excluded features with poor sta -\nbility and consistency, there might still be some outlining errors.\n5.  Conclusion\nThe findings showed that a combined model based on \nnon-enhanced MRI radiomics was effective in predicting the \nefficacy of HIFU ablation of adenomyosis before surgery. This \nmodel will assist clinicians in treatment decisions and the \nscreening of patients who are likely to benefit from HIFU.\nAuthor contributions\nZiyi Liu and Ziyan Liu contributed equally to this work. (1) Conception \nand design: Ziyi Liu, Ziyan Liu and Xiaohua Huang. (2) Collection and \nassembly of data: Xiyao Wan and Yuan Wang. (3) Data analysis: Ziyi Liu \nand Ziyan Liu. (4) Manuscript writing: Ziyi Liu. (5) Final approval of man -\nuscript: All authors.\nEthical approval\nThe Affiliated Hospital of North Sichuan Medical College’s Medical Ethics \nCommittee eliminated the need for informed consent after approving \nthe research plan (IRB no.2024ER282-1).\nDisclosure statement\nNo potential conflict of interest was reported by the author(s).\nFunding\nThis study has received funding by Nanchong City School Cooperation \nProject [No. 19SXHZ0429] and This work was supported by Bureau of \nScience and Technology Nanchong Municipality.\nFigure 4.  (a ) Decision curves of clinical-imaging model, radiomics model, and combined model; (B) Combined model calibration curves in the training set and \ntest set.\n\nINTERNATIONAL  JOURNAL  OF HYPERTHERMIA 7\nORCID\nXiaohua Huang  http://orcid.org/0000-0002-3490-4142\nData availability statement\nThe data used to support the findings of this study are available from \nthe corresponding author upon request. The data are not publicly avail -\nable because they contain information that can compromise the privacy \nof the research participants.\nReferences\n [ 1] Jin W, Wang S, Wang T, et  al. Multi-machine learning model based \non habitat subregions for outcome prediction in adenomyosis \ntreated by uterine artery embolization. Acad Radiol. 2024;31(12):  \n4985–4995. doi: 10.1016/j.acra.2024.05.037.\n [ 2] Li Z, Zhang J, Song Y, et  al. Utilization of radiomics to predict \nlong-term outcome of magnetic resonance-guided focused ultra -\nsound ablation therapy in adenomyosis. Eur Radiol. 2021;31(1):392–\n402. doi: 10.1007/s00330-020-07076-1.\n [ 3] Bae JS, Lee JY, Chung HH, et  al. Optimized treatment parameter by \ncomputer simulation for high-intensity focused ultrasound treat -\nment of uterine adenomyosis: short-term and long-term results. \nPLoS One. 2024;19(3):e0301193. doi: 10.1371/journal.pone.0301193.\n [ 4] Loring M, Chen TY, Isaacson KB. A systematic review of adenomyo -\nsis: it is time to reassess what we thought we knew about the \n disease. J Minim Invasive Gynecol. 2021;28(3):644–655. doi: 10.1016/ \nj.jmig.2020.10.012.\n [ 5] Tan J, Yong P , Bedaiwy MA. A critical review of recent advances in \nthe diagnosis, classification, and management of uterine adenomy -\nosis. Curr Opin Obstet Gynecol. 2019;31(4):212–221. doi: 10.1097/\ngco.0000000000000555.\n [ 6] Du CC, Wang YQ, Qu DC, et  al. Magnetic resonance imaging T2WI \nhyperintense foci number and the prognosis of adenomyosis after \nhigh-intensity focused ultrasound treatment. Int J Gynaecol Obstet. \n2021;154(2):241–247. doi: 10.1002/ijgo.13587.\n [ 7] Lee JS, Hong GY, Lee KH, et  al. Safety and efficacy of \nultrasound-guided high-intensity focused ultrasound treatment for \nuterine fibroids and adenomyosis. Ultrasound Med Biol. 2019;45(12):  \n3214–3221. doi: 10.1016/j.ultrasmedbio.2019.08.022.\n [ 8] Liu Y, Zhang WW, He M, et  al. Adverse effect analysis of \nhigh-intensity focused ultrasound in the treatment of benign uter -\nine diseases. Int J Hyperthermia. 2018;35(1):56–61. doi: 10.1080/ \n02656736.2018.1473894.\n [ 9] Yao R, Hu J, Zhao W, et  al. A review of high-intensity focused \nultrasound as a novel and non-invasive interventional radiology \ntechnique. J Interv Med. 2022;5(3):127–132. doi: 10.1016/j.jimed.  \n2022.06.004 .\n [10] Guo Q, Xu F, Ding Z, et  al. High intensity focused ultrasound treat -\nment of adenomyosis: a comparative study. Int J Hyperthermia. \n2018;35(1):505–509. doi: 10.1080/02656736.2018.1509238.\n [11] Mindjuk I, Trumm CG, Herzog P , et  al. MRI predictors of clinical \nsuccess in MR-guided focused ultrasound (MRgFUS) treatments of \nuterine fibroids: results from a single centre. Eur Radiol. \n2015;25(5):1317–1328. doi: 10.1007/s00330-014-3538-6.\n [12] Si M, Lv F, Tang M, et  al. Non-contrast enhanced MRI for efficiency \nevaluation of high-intensity focused ultrasound in adenomyosis \nablation. Int J Hyperthermia. 2024;41(1):2295813. doi: 10.1080/ \n02656736.2023.2295813.\n [13] Fan TY, Zhang L, Chen W, et  al. Feasibility of MRI-guided high in -\ntensity focused ultrasound treatment for adenomyosis. Eur J Radiol. \n2012;81(11):3624–3630. doi: 10.1016/j.ejrad.2011.05.036.\n [14] Lee JS, Hong GY, Park BJ, et  al. Ultrasound-guided high-intensity \nfocused ultrasound treatment for uterine fibroid & adenomyosis: a \nsingle center experience from the Republic of Korea. Ultrason \nSonochem. 2015;27:682–687. doi: 10.1016/j.ultsonch.2015.05.033.\n [15] Shui L, Mao S, Wu Q, et  al. High-intensity focused ultrasound \n(HIFU) for adenomyosis: two-year follow-up results. Ultrason \nSonochem. 2015;27:677–681. doi: 10.1016/j.ultsonch.2015.05.024.\n [16] Bazot M, Daraï E. Role of transvaginal sonography and magnetic \nresonance imaging in the diagnosis of uterine adenomyosis. Fertil \nSteril. 2018;109(3):389–397. doi: 10.1016/j.fertnstert.2018.01.024.\n [17] Sudderuddin S, Helbren E, Telesca M, et  al. MRI appearances of be -\nnign uterine disease. Clin Radiol. 2014;69(11):1095–1104. doi: \n10.1016/j.crad.2014.05.108.\n [18] Gong C, Setzen R, Liu Z, et  al. High intensity focused ultrasound \ntreatment of adenomyosis: the relationship between the features \nof magnetic resonance imaging on T2 weighted images and the \ntherapeutic efficacy. Eur J Radiol. 2017;89:117–122. doi: 10.1016/j.\nejrad.2017.02.001.\n [19] Yu JW, Yang MJ, Jiang L, et  al. Factors influencing USgHIFU abla -\ntion for adenomyosis with NPVR ≥ 50. Int J Hyperthermia. \n2023;40(1):2211753. doi: 10.1080/02656736.2023.2211753.\n [20] Lafata KJ, Wang Y, Konkel B, et  al. Radiomics: a primer on \nhigh-throughput image phenotyping. Abdom Radiol (NY). \n2022;47(9):2986–3002. doi: 10.1007/s00261-021-03254-x.\n [21] Mayerhoefer ME, Materka A, Langs G, et  al. Introduction to radio -\nmics. J Nucl Med. 2020;61(4):488–495. doi: 10.2967/jnumed. \n118.222893.\n [22] Ying J, Jing X, Gao F, et  al. Prediction of ablation rate for \nhigh-intensity focused ultrasound therapy of adenomyosis in MR \nimages based on multi-model fusion. J Imaging Inform Med. \n2024;37(4):1579–1590. doi: 10.1007/s10278-024-01063-4.\n [23] Jha RC, Zanello PA, Ascher SM, et  al. Diffusion-weighted imaging \n(DWI) of adenomyosis and fibroids of the uterus. Abdom Imaging. \n2014;39(3):562–569. doi: 10.1007/s00261-014-0095-z.\n [ 24] Keserci B, Duc NM. Magnetic resonance imaging features influencing \nhigh-intensity focused ultrasound ablation of adenomyosis with a non-\nperfused volume ratio of ≥90% as a measure of clinical treatment suc-\ncess: retrospective multivariate analysis. Int J Hyperthermia. 2018;35(1): \n626–636. doi: 10.1080/02656736.2018.1516301.\n [25] Zhou M, Chen JY, Tang LD, et  al. Ultrasound-guided high-intensity \nfocused ultrasound ablation for adenomyosis: the clinical experi -\nence of a single center. Fertil Steril. 2011;95(3):900–905. doi: \n10.1016/j.fertnstert.2010.10.020.\n [26] Lambin P , Leijenaar RTH, Deist TM, et  al. Radiomics: the bridge be -\ntween medical imaging and personalized medicine. Nat Rev Clin \nOncol. 2017;14(12):749–762. doi: 10.1038/nrclinonc.2017.141.\n [27] Rizzo S, Botta F, Raimondi S, et  al. Radiomics: the facts and the \nchallenges of image analysis. Eur Radiol Exp. 2018;2(1):36. doi: \n10.1186/s41747-018-0068-z.\n [28] Zhou J, Yu X, Wu Q, et  al. Radiomics analysis of intratumoral and \ndifferent peritumoral regions from multiparametric MRI for evaluat -\ning HER2 status of breast cancer: a comparative study. Heliyon. \n2024;10(7):e28722. doi: 10.1016/j.heliyon.2024.e28722.\n [29] Burla L, Sartoretti E, Mannil M, et  al. MRI-based radiomics as a \npromising noninvasive diagnostic technique for adenomyosis.  \nJ Clin Med. 2024;13(8):2344. doi: 10.3390/jcm13082344.\n [30] Luo Y, Sun X, Kong X, et  al. A DWI-based radiomics-clinical machine \nlearning model to preoperatively predict the futile recanalization af -\nter endovascular treatment of acute basilar artery occlusion patients. \nEur J Radiol. 2023;161(2023):110731. doi: 10.1016/j.ejrad.2023.110731.\n [31] Shih IL, Yen RF, Chen CA, et  al. PET/MRI in cervical cancer: associ -\nations between imaging biomarkers and tumor stage, disease pro -\ngression, and overall survival. J Magn Reson Imaging. \n2021;53(1):305–318. doi: 10.1002/jmri.27311.\n [ 32] Gong C, Yang B, Shi Y, et  al. Factors influencing the ablative efficiency \nof high intensity focused ultrasound (HIFU) treatment for adenomyo-\nsis: a retrospective study. Int J Hyperthermia. 2016;32(5):496–503. doi: \n10.3109/02656736.2016.1149232.\n [33] Qin S, Jiang Y, Wang F, et  al. Development and validation of a \ncombined model based on dual-sequence MRI radiomics for \n predicting the efficacy of high-intensity focused ultrasound abla -\ntion for hysteromyoma. Int J Hyperthermia. 2023;40(1):2149862. \ndoi: 10.1080/02656736.2022.2149862.","source_license":"CC0","license_restricted":false}