{"paper_id":"8fa7659c-b843-48e1-80ed-43e88e1aa561","body_text":"Multi-scale radiomics in pelvic MRI for\nendometriosis subtyping: highlighting data\nheterogeneity constraints\nEliot Leguy1,2, Chloe Mallet2, Nicolas Normand2, and Elodie Germani1\n1 Laboratoire Traitement du Signal et de l’Image (LTSI, INSERM UMR 1099),\nUniversité de Rennes, Rennes, France\n2 Nantes Université, Ecole Centrale Nantes, CNRS, LS2N, UMR 6004, Nantes,\nFrance\nCorresponding author:elodie.germani@univ-rennes.fr\nAbstract.Analyzing female pelvic MRIs is challenging, especially for\nevaluating endometriosis, where visual features are influenced by sev-\neral factors, including anatomical complexity, technical variability, and\ninter-reader variability. Here, we evaluate a radiomics-based pipeline for\npatient-level endometriosis subtyping using the publicly available UT-\nEndoMRI dataset. We extract radiomics features from manually seg-\nmented uterine and ovarian regions and compare several multi-scale fea-\nture representations and feature-selection strategies. We train supervised\nclassifiers to distinguish patients with at least one endometrioma from\nthose without, and perform an unsupervised perturbation analysis to\nassess whether radiomics profiles reveal reproducible patient subgroups.\nThe best supervised performance is achieved using raw Wavelet-derived\nfeatures and a Gradient Boosting classifier, yielding an AUC of 0.80.\nHowever, this model produces several false positives, resulting in low\nspecificity. ComBat harmonization does not consistently improve perfor-\nmance, suggesting that post hoc harmonization is insufficient in a small,\nmulti-site cohort in which acquisition groups contained very few patients.\nUsing an unsupervised clustering analysis, we identify reproducible but\npoorly separated partitions that remain associated with acquisition vari-\nables. Overall, these results suggest that pelvic MRI radiomics contain\na preliminary signal for endometriosis subtyping, while highlighting the\nfragility of radiomics-based subtyping in small, multi-site datasets.\nKeywords:Radiomics·Endometriosis·Subtyping·MRI·Multi-site\n1 Introduction\nEndometriosis is a chronic inflammatory disease affecting around 10% of women\nworldwide and characterized by the presence of endometrial-like tissue (also\ncalled “lesions”) outside the uterus. Disease diagnosis and treatment planning\nrequire detailed evaluation of lesions, including their location, extension, and\ntissue composition [3]. To this end, pelvic magnetic resonance imaging (MRI) is\narXiv:2607.26692v1  [eess.IV]  29 Jul 2026\n\n2 Leguy, Mallet et al.\na central tool for a non-invasive evaluation of endometriosis [1]. Recently, scor-\ning systems such as the deep pelvic endometriosis index (dPEI) [13] have helped\nstandardize this evaluation, providing a reproducible pipeline for characterizing\nendometriosis status and facilitating treatment planning. However, this scoring\nsystem relies on visual examination of pelvic MRI to locate lesions and measure\ntheir extent. However, due to the complexity of the female pelvic anatomy, the\nimpact of acquisition parameters on image appearance, and the level of expertise\nrequired to identify and locate endometriosis lesions, this process remains chal-\nlenging and prone to inter-reader variability, especially for junior readers [11].\nInthiscontext,radiomicscanofferacomplementaryquantitativeapproachto\nvisual examination by extracting structured image descriptors, including shape,\nintensity, and texture features, from regions of interest [4]. Multi-scale filtering,\nsuch as Wavelet decomposition and Laplacian of Gaussian (LoG) filtering, can\nfurther characterize spatial patterns at different frequencies. In female pelvic\nimaging, radiomics-based methods have shown promising results for assessing\nadenomyosis [2], classifying endometrial lesions [9], and differentiating ovarian\nendometriomas from ovarian dermoid cysts [10].\nHowever, radiomics features are sensitive to several sources of variability,\nsuch as technical variability, in which scanner differences, acquisition protocols,\nreconstruction algorithms, and image resolution can alter feature distributions\nindependentlyofpathology.Inmulti-sitestudies,harmonizationmethodssuchas\nComBat [12] have been proposed to reduce technical variability while preserving\nclinically relevant covariates. In addition, standard radiomics feature-extraction\npipelines produce a large number of non-independent features that must be\nfiltered for subsequent use. Feature selection methods such as LASSO [14] and\nBoruta [7] can help limit the effects of high dimensionality in small cohorts and\nevaluate residual technically related bias.\nHere, we evaluate the potential and challenges of radiomics for endometrio-\nsis subtyping from a publicly available multi-centric pelvic MRI dataset [8].\nOur contributions are threefold. First, we develop a pipeline for multi-scale ra-\ndiomics feature extraction from organ segmentations in T1w and T2w pelvic\nMRIs and train machine-learning classifiers to distinguish patients with and\nwithout endometrioma. Second, we explore the effects of inter-site harmoniza-\ntion on performance using ComBat in settings with small sample sizes per site\nand unknown covariate effects. Third, we perform an unsupervised analysis of\nradiomics perturbation profiles to investigate whether the feature space reveals\nclinically meaningful subgroups or is primarily driven by technical variability.\n2 Methods\n2.1 Radiomics feature extraction pipeline\nLetI i,s denote the MRI three-dimensional (3D) volume of patientifor sequence\ns, and letΩi,r ⊂Z 3 denote the segmented region of interest (ROI) corresponding\nto anatomical regionr. Radiomics extraction is defined as the mapping\nΦ: (I i,s, Ωi,r)→x i,s,r ∈R p,\n\nMRI Radiomics for Endometriosis Subtyping 3\nwherex i,s,r is a vector of image-derived descriptors.\nTo assess the contribution of multi-scale information, radiomics features are\nextracted from the original and filtered images. Wavelet filtering decomposes\neach image into eight 3D sub-bands(LLL, LLH, . . . , HHH), whereLandH\ndenote low-pass and high-pass filtering along each spatial axis. These decom-\npositions capture complementary low-frequency anatomical structure and high-\nfrequency local texture. LoG filtering is used to emphasize local intensity tran-\nsitions at multiple spatial scales. For an imageI, the LoG response at scaleσis\ndefined asL σ =∇ 2(Gσ ⊛I),whereG σ is a Gaussian kernel.\n2.2 Inter-site harmonization\nTo reduce the effects of technical variability on radiomics features, we explore\nthe benefits of ComBat harmonization [12]. For each radiomics feature, ComBat\nmodels the observed value as:\nyij =α+x ⊤\nijβ+γ i +δ iεij,\nwherey ij isthevalueofthefeaturemeasuredforsubjectorregionjinacquisition\nsettingi,αis the global feature mean,x ⊤\nijβrepresents the effect of preserved\ncovariates,γ i is an additive batch effect,δi is a multiplicative batch effect, andεij\nis the residual error. In our case,x⊤\nijβrepresents the expected contribution of the\nendometrioma label to the radiomics feature value, so that this disease-related\ninformation is preserved while correcting for acquisition-related effects.\nAfter estimating the batch parameters, the corrected feature value are:\nyComBat\nij = yij −ˆα−x ⊤\nij ˆβ−ˆγi\nˆδi\n+ ˆα+x⊤\nij ˆβ,\nso that site-related additive and multiplicative effects are reduced while the\nspecified covariate effects are preserved. Although ComBat is widely used for\nimaging biomarker harmonization, it depends on several assumptions. Reliable\nestimation of site effects requires sufficient samples per acquisition site and bal-\nanced covariate distributions. Previous work [12,6] has recommended between\n20 and 30 subjects per site, while more recent studies suggest that reasonable\nharmonization can be obtained with 16 to 32 subjects in a moving site when\na well-populated reference site is available. ComBat also assumes that covari-\nate effects are comparable across sites. Violations of this assumption, or strong\nconfounding between site and disease, can lead to erroneous harmonization or\nremoval of clinically relevant signals. Here, we analyze ComBat-harmonized fea-\ntures alongside raw features rather than using them as a guaranteed correction.\n2.3 Feature selection\nRadiomics features extraction produces a high-dimensional feature space with\nsubstantial redundancy between descriptors. To address these challenges, we per-\nformatwo-stepfeatureselectionstrategy.First,weuseLASSOlogisticregression\n\n4 Leguy, Mallet et al.\nto identify sparse linear predictors, retaining only features with non-zero coeffi-\ncients. Second, we use Boruta [7] to identify features with non-linear predictive\nrelevance. This method trains a Random Forest classifier on the original features\naugmented with randomized shadow features. A feature is retained only if its\nimportance is consistently higher than the importance of randomized shadow\nfeatures. For each configuration, the final signature is defined as the intersection\nof the LASSO-selected and Boruta-selected features.\n2.4 Supervised classification\nThe supervised task is formulated as a binary classification problem. For each\npatient-level observation, the label is defined asyi ∈ {0,1},wherey i = 1denotes\na patient with at least one confirmed endometrioma andyi = 0denotes a pa-\ntient without any identified endometrioma. The objective is to evaluate whether\nradiomics features extracted from pelvic MRI can discriminate endometrioma-\npositive from endometrioma-negative patients. We evaluate four candidate clas-\nsifiers: Logistic Regression, Random Forest Classifier with 100 trees, Support\nVector Machine with radial basis function kernel, and Gradient Boosting classi-\nfier. Each model is embedded in a feature standardization pipeline. Candidate\nmodels are compared using five-fold cross-validation on the training set, with\nthe area under the receiver operating characteristic curve (AUC) as the crite-\nrion for selecting the final model. The selected model is retrained on the whole\ntraining set and evaluated on the held-out test set. Evaluation metrics include\nAUC, accuracy, sensitivity, specificity, precision, and F1-score.\n2.5 Unsupervised Perturbation Analysis\nTo further assess the discriminative power of radiomics features, we perform an\nunsupervised evaluation of patient-level radiomics profiles. We define the refer-\nence control population as a sample of 8 patients, identified from the cohort data\nas having neither endometriosis nor endometrioma. Because the dataset reports\nthe number of patients without endometriosis but does not explicitly provide\ntheir identifiers, this control group was derived from the available metadata and\nshould be interpreted cautiously. For each feature, patient values were converted\ninto robust perturbation scores relative to this control population:\nzi,j = xi,j −median(X control,j)\nIQR(Xcontrol,j) ,\nwhereX control,j denotes the distribution of featurejamong the inferred control\npatients.Featureswithzerointerquartilerangeinthecontrolgroupareexcluded.\nWe then perform dimensionality reduction using Principal Component Anal-\nysis (PCA), retaining the smallest number of components explaining 90% of\nthe total variance. We apply K-means clustering to these components fork∈\n{2, . . . ,6}, and select the optimal number of clusters by maximizing the silhou-\nette coefficient. Cluster stability is assessed using 50 bootstrap samples and the\n\nMRI Radiomics for Endometriosis Subtyping 5\nCharacteristic V alue\nTotal patients 51\nEndometrioma-positive patients 40\nEndometrioma-negative patients 11\nPrimary MRI sequences T2, T1FS\nAnalyzed anatomical regions Ovary, uterus\n(a) Dataset characteristics.\n (b) Left: T1 FatSat, Right: T2 slices.\nFig.1: Summary of the UT-EndoMRI D1 cohort.\nAdjusted Rand Index (ARI). To evaluate whether the resulting clusters reflect\nclinical structure or residual acquisition bias, associations between cluster labels\nand technical variables, such as imaging site and scanner model, are quantified\nusingχ 2 tests and Cramer’sV.\n3 Experiments\nDatasetWe conduct our experiments on the UTHealth Endometriosis MRI\ndataset (UT-EndoMRI) [8], a publicly available dataset containing pelvic multi-\nsequence MRI and manual organs and lesions segmentations from women with\nendometriosis. The complete dataset is organized into two cohorts. Here, we\nuse the first cohort, denoted D1, which contains 51 patients as described in\nTab 1a, acquired before 2022 from the Memorial Hermann Hospital System and\nthe Texas Children’s Hospital Pavilion for Women. We use T2- and T1-weighted\nfat-suppressed (T1FS) MRI sequences and manual segmentations of the uterus\nand ovaries. Radiomics features from both organs are aggregated at the patient\nlevel. This dataset comprises 51 patients, of whom 40 are endometrioma-positive,\nand 11 are endometrioma-negative. We construct a balanced held-out test set\ncontaining 5 positive and 5 negative patients. The remaining patients (N=41, 35\npositive and 6 negative) are used for training and cross-validation.\nTab. 1a summarizes the dataset used for the classification experiments.\nImplementation detailsIn the supervised classification task, we compare\nmodels trained with four radiomics configurations: Wavelet, Wavelet+ComBat,\nLoG, and LoG+ComBat. Features are extracted using PyRadiomics v3.0.1 [4],\nfollowing the Image Biomarker Standardization Initiative (IBSI) [5]. The ex-\ntracted features include shapes, first-order statistics, GLCM, GLRLM, GLSZM,\nand GLDM descriptors. MRI volumes are normalized to a scale of 100 and re-\nsampled to an isotropic1.0mm spacing using B-spline interpolation. All prepro-\ncessing and feature extraction are performed in the volume space. In the LoG\nconfiguration, Gaussian scales are set toσ∈ {1.0,2.0,3.0}mm. When multi-\nple segmentation masks are available for the same patient, organ, and sequence,\nfeature values are averaged to produce a single patient-level representation for\neach configuration. Missing or non-finite values are handled by mean imputation\nfitted on the full training set, followed by training-fitted z-score standardization.\n\n6 Leguy, Mallet et al.\nTable 1: Performance of radiomics classifiers for endometriosis subtyping. CV\nAUC is reported as mean±standard deviation across cross-validation folds.\nOther metrics are computed on the held-out test set.\nConfiguration Model CV AUC Test AUCAcc.Sens. Spec. Prec. F1\nWavelet raw GB 0.936±0.108 0.800 0.60 1.000 0.200 0.556 0.714\nWavelet ComBat LR 0.955±0.065 0.680 0.60 0.800 0.400 0.571 0.667\nLoG raw LR 0.864±0.135 0.520 0.50 0.400 0.600 0.500 0.444\nLoG ComBat RF 0.929±0.101 0.580 0.60 0.800 0.400 0.571 0.667\nTo prevent leakage, ComBat parameters are estimated on the training set only\nand then applied to the held-out test set. No test-set information is used to\nestimate these harmonization parameters. Feature selection is performed on the\nfull train set once and kept fixed for the rest of the experiments. For classifier\nselection, we use cross-validation on this fixed training-derived set of features,\nwhile final performance is assessed on the held-out test set. Feature selection\nis not repeated inside each cross-validation fold. Therefore, cross-validation is\nused to optimize the classifier on a fixed training-derived signature, while final\ngeneralization is assessed only on the independent held-out test set.\n3.1 Supervised Classification Results\nTab. 1 summarizes the performance of the four radiomics configurations. The\nhighest cross-validation AUC is obtained with ComBat-harmonized Wavelet fea-\ntures and Logistic Regression (AUC=0.955±0.065). However, the highest held-\nout test AUC is obtained with raw Wavelet features and Gradient Boosting\n(AUC=0.80). With this configuration, the model detects positive cases, but its\nlow specificity also indicates many false positives. Given the 10-patient test set,\nthe difference between raw and ComBat-harmonized Wavelet features should\nnot be overinterpreted. These results do not show a clear benefit of ComBat\nin this cohort, possibly because site- and scanner-specific effects are estimated\nfrom small and imbalanced acquisition groups. LoG-based models perform worse\nthan Wavelet-based models, with test AUCs of 0.52 without harmonization and\n0.58 after ComBat. Overall, these results suggest preliminary radiomics signals\nfor endometriosis subtyping in Wavelet-derived texture features of uterine and\novarian masks. However, the discrepancy between cross-validation and held-out\nperformance shows that the models’ performance remains unstable. Given the\nsmall size of the test set, these results should be interpreted as exploratory rather\nthan as evidence of clinical generalization.\n3.2 Radiomic feature analysis\nThe selected features include shape, first-order, and texture descriptors, sug-\ngesting that the classification signal is driven not only by global morphology\nbut also by intensity distribution and local spatial heterogeneity. In the Wavelet\n\nMRI Radiomics for Endometriosis Subtyping 7\nRaw features ComBat harmonized features\nFig.2: Projection of the first two components from the PCA of radiomics pertur-\nbation profiles after K-means clustering withk= 2. The control patients (circled\ndots) have no endometriomas. Dots’ colors represent their acquisition site.\nTable 2: Quality and technical association of the unsupervised clustering results:\nbest silhouette score, Bootstrap stability measured by ARI, and the association\nwith acquisition site usingχ2 p-values and Cramer’sV.\nConfiguration k Silh. ARI K-means Site(p;V) Scanner(p;V)\nRaw, all organs 2 0.371 0.859±0.095 0.461; 0.569 0.386; 0.417\nComBat, all organs2 0.318 0.900±0.087 0.106; 0.616 0.310; 0.417\nRaw, ovary 2 0.241 0.699±0.154 0.200; 0.652 0.186; 0.560\nRaw, uterus 2 0.278 0.806±0.177 0.118; 0.649 0.665; 0.467\nComBat, ovary 2 0.330 0.879±0.168 0.022; 0.768 0.528; 0.480\nComBat, uterus 2 0.263 0.769±0.276 0.412; 0.529 0.164; 0.680\nconfigurations, selected descriptors are mainly derived from high-frequency sub-\nbands, indicating that directional multi-scale texture information contributes\nto the supervised signal observed in the classification experiments. LoG-derived\nsignatures retained features are related to local intensity transitions, but the\ncorresponding models show weaker performance.\n3.3 Unsupervised clustering of radiomics perturbation profiles\nTo investigate whether the radiomics feature space contains clinically relevant\npatient subgroups, we perform an unsupervised clustering analysis based on\nperturbation profiles. Each patient is represented by a z-score vector computed\nrelative to a control population without endometriosis or endometrioma.\nThe clustering analysis does not seem to reveal distinct clinical phenotypes.\nAcrossallconfigurations,thesilhouettecriterionisoptimalfork= 2,butthecor-\nresponding silhouette scores remain modest, ranging from0.241to0.371. Since\nclearly separated clusters are typically expected to have substantially higher sil-\nhouette scores, often above0.7, these results suggest weak separation and should\nnot be interpreted as evidence of robust clinical subtypes. The PCA projections\nin Fig. 2 show substantial overlap between the two groups. Together with the\nmodest silhouette scores reported in Tab. 2, this suggests that the dominant\n\n8 Leguy, Mallet et al.\nradiomic structure corresponds to a broad perturbation gradient rather than to\nwell-separated patient subtypes. Despite this limited separation, the two-cluster\npartitions are relatively stable under resampling. Bootstrap ARI values for K-\nmeans range from0.699±0.154to0.900±0.087, indicating moderate to high\nclustering stability. This combination of modest silhouette scores and stable ARI\nsuggests that the feature space contains a reproducible partition, but not one\nthat should automatically be interpreted as clinically meaningful. The associa-\ntion analysis further supports this cautious interpretation. Several configurations\nshow high Cramer’sVvalues for the acquisition variables, even when the cor-\nrespondingχ 2 tests are not statistically significant, likely due to the limited\ncohort size. The strongest site association is observed after harmonization in the\novary-specific analysis (p= 0.022,V= 0.768), indicating that ComBat does\nnot fully remove technical structure from the radiomics space. Therefore, the\nclustering results suggest a residual acquisition signature rather than evidence\nof reproducible endometriosis subtypes.\n4 Conclusion\nThis study evaluates a radiomics-based pipeline for patient-level endometriosis\nsubtyping from pelvic MRI. We compare multi-scale radiomics features based\non Wavelet and LoG filtering, with and without ComBat harmonization, and\nobtain the best held-out performance with raw Wavelet features, reaching a test\nAUC of 0.80 on a balanced test set, suggesting that Wavelet-derived texture\nfeatures might contain discriminative information for endometriosis subtyping.\nHowever, cross-validation performance is consistently higher than held-out\nperformance, suggesting model instability in the small-sample setting. Inter-site\nharmonization with ComBat does not improve generalization: it reduces the test\nAUC for Wavelet features and yields only a modest improvement for LoG fea-\ntures. This observation is likely related to the limited number of patients per\nacquisition site or scanner group, often ranging from 1 to 10, which is below the\ncommonly recommended sample sizes for reliable ComBat estimation. Cluster-\ning of radiomics perturbation profiles does not reveal clearly separated clinical\nphenotypes. Instead, clusters are strongly associated with acquisition variables\nsuch as site and scanner, indicating that a substantial part of the radiomics\nstructure reflects technical variability rather than disease-specific subtypes.\nOverall, this work provides preliminary evidence that multi-scale radiomics\nfeatures, particularly Wavelet-derived descriptors, can contribute to MRI-based\nendometriosissubtyping.Atthesametime,ithighlightsthefragilityofradiomics\nmodels in small multi-center cohorts and the practical limits of post hoc har-\nmonization when site-level sample sizes are small. Future work should focus\non larger external validation cohorts, stricter control of acquisition protocols,\nbetter-balanced site distributions, and more robust domain-adaptation or har-\nmonization strategies before such models can be considered clinically reliable.\n\nMRI Radiomics for Endometriosis Subtyping 9\nReferences\n1. Bazot, M., Darai, E., Hourani, R., Thomassin, I., Cortez, A., Uzan,\nS., Buy, J.N.: Deep pelvic endometriosis: MR imaging for diagnosis and\nprediction of extension of disease. Radiology232(2), 379–389 (2004).\nhttps://doi.org/10.1148/radiol.2322030762\n2. Burla, L., Sartoretti, E., Mannil, M., Seidel, S., Sartoretti, T., Krentel, H.,\nDe Wilde, R.L., Imesch, P.: Mri-based radiomics as a promising noninvasive di-\nagnostic technique for adenomyosis. Journal of Clinical Medicine13(8) (2024).\nhttps://doi.org/10.3390/jcm13082344\n3. Chapron, C., Marcellin, L., Borghese, B., Santulli, P.: Rethinking mechanisms, di-\nagnosis and management of endometriosis. Nature Reviews Endocrinology15(11),\n666–682 (2019). https://doi.org/10.1038/s41574-019-0245-z\n4. van Griethuysen, J.J., Fedorov, A., Parmar, C., Hosny, A., Aucoin, N., Narayan,\nV., Beets-Tan, R.G., Fillion-Robin, J.C., Pieper, S., Aerts, H.J.: Computational\nradiomics system to decode the radiographic phenotype. Cancer Research77(21),\ne104–e107 (10 2017). https://doi.org/10.1158/0008-5472.CAN-17-0339\n5. Image Biomarker Standardisation Initiative: Ibsi reference manual and guidelines.\nhttps://theibsi.github.io/ (2026), last accessed: 2026-04-24\n6. Jodoin, P.M., Edde, M., Girard, G., Dumais, F., Theaud, G., Dumont, M.,\nHoude, J.C., David, Y., Descoteaux, M., Initiative, A.D.N.: Challenges and\nbest practices when using ComBAT to harmonize diffusion MRI data. Scien-\ntific Reports15(1), 41508 (2025). https://doi.org/10.1038/s41598-025-25400-x,\nhttps://doi.org/10.1038/s41598-025-25400-x\n7. Kursa, M.B., Jankowski, A., Rudnicki, W.R.: Boruta: A system for\nfeature selection. Fundamenta Informaticae101(4), 271–285 (2010).\nhttps://doi.org/10.3233/FI-2010-288\n8. Liang, X., Alpuing Radilla, L.A., Khalaj, K., Dawoodally, H., Mokashi,\nC., Guan, X., Roberts, K.E., Sheth, S.A., Tammisetti, V.S., Gian-\ncardo, L.: Uthealth - endometriosis mri dataset (ut-endomri) (Jun 2025).\nhttps://doi.org/10.5281/zenodo.15750762\n9. Liu, J., Lin, H., Pang, P., Luo, P., Fan, B., Yu, J.: Development of MRI-based\nradiomics predictive model for classifying endometrial lesions. Scientific Reports\n13(1), 1590 (2023). https://doi.org/10.1038/s41598-023-28819-2\n10. Liu, L., Cai, W., Zhou, C., Tian, H., Wu, B., Zhang, J., Yue, G., Hao, Y.: Ul-\ntrasound radiomics-based artificial intelligence model to assist in the differential\ndiagnosis of ovarian endometrioma and ovarian dermoid cyst. Frontiers in Medicine\nVolume 11 - 2024(2024). https://doi.org/10.3389/fmed.2024.1362588\n11. Miceli, V., Ninkova, R.V., Gennarini, M., Riccardi, S., Curti, F., Amoroso, C.,\nManghi,R.,Porpora,M.G.,Thomassin-Naggara,I.,Manganaro,L.:Structuredmri\nreporting in patients with deep pelvic endometriosis: performance and clinical im-\npact of the deep pelvic endometriosis index (dpei). European Journal of Radiology\n202, 112926 (2026). https://doi.org/https://doi.org/10.1016/j.ejrad.2026.112926\n12. Orlhac,F.,Eertink,J.J.,Cottereau,A.S.,Zijlstra,J.M.,Thieblemont,C.,Meignan,\nM., Boellaard, R., Buvat, I.: A guide to combat harmonization of imaging biomark-\ners in multicenter studies. Journal of Nuclear Medicine63(2), 172–179 (2022).\nhttps://doi.org/10.2967/jnumed.121.262464\n13. Thomassin-Naggara, I., Lamrissi, A., Sorbets, E., Darai, E., Rouzier, R., Bazot,\nM.:Magneticresonanceimagingclassificationofdeeppelvicendometriosis:descrip-\ntion and impact on surgical management. Human Reproduction35(5), 1098–1108\n(2020). https://doi.org/10.1093/humrep/deaa103\n\n10 Leguy, Mallet et al.\n14. Tibshirani, R.: Regression shrinkage and selection via the lasso. Journal of the\nRoyal Statistical Society: Series B (Methodological)58(1), 267–288 (01 1996).\nhttps://doi.org/10.1111/j.2517-6161.1996.tb02080.x","source_license":"CC0","license_restricted":false}