{"paper_id":"f5a68304-9dad-4fac-9ca0-d9da64d8564f","body_text":"117\nCopyright © 2025 Sungkyunkwan University School of Medicine\nThis is an Open Access article \ndistributed under the terms of the \nCreative Commons Attribution \nNon-Commercial License (https://\ncreativecommons.org/licenses/\nby-nc/4.0/).\nREVIEW \nARTICLE\nMachine learning in the early detection of \nendometriosis: a literature review on symptom \nclustering and imaging integration \nA literature review\nZiad Mumtaz Ramadan\n1\n, Saaid Mounzer Mouazen\n1\n, Shady Saud Khan\n1\n,  \nSariya Khan\n1\n, Noha S. Farag\n2\n1\nGeneral Medicine Practice Program, Batterjee Medical College, Jeddah, Saudi Arabia\n2\nDepartment of Clinical Pathology, Batterjee Medical College, Jeddah, Saudi Arabia\nABSTRACT\nEndometriosis is a gynecologic inflammatory condition that affects up to 10% of repro-\nductive-aged women worldwide. The disease exhibits heterogeneous presentations \nand is associated with a prolonged diagnostic delay, often exceeding seven years, be-\ncause existing diagnostic modalities such as transvaginal ultrasound, magnetic reso-\nnance imaging, and the biomarker cancer antigen 125 (CA-125) are suboptimal. This re-\nview examines how machine learning (ML) is playing an increasingly significant role in \nearly, non-surgical endometriosis diagnosis through two main approaches: symptom \nclustering and imaging integration. Unsupervised ML algorithms such as k-means, par-\ntitioning around medoids, and Bayesian networks have demonstrated success in iden-\ntifying clinically informative endometriosis phenotypes from patient-reported symp-\ntoms and electronic health records. Concurrently, ML models such as convolutional \nneural networks and radiomics approaches have achieved high accuracy in lesion de-\ntection from imaging data, in some cases surpassing human interpretation. Despite \nthese advances, significant challenges remain, including limited access to large, anno-\ntated multimodal datasets, the absence of widely accepted evaluation standards, and \nconcerns regarding interpretability and generalizability. Multicenter, integrative studies \nand the incorporation of explainability techniques are recommended as potential strat-\negies to address these gaps. Finally, multimodal ML approaches that combine symp-\ntomatology and imaging data hold substantial promise for reducing diagnostic delays, \nfacilitating early intervention, and improving clinical outcomes in the management of \nendometriosis.\nKeywords: Artificial intelligence; Endometriosis; Machine learning\nPrecision and Future Medicine 2025;9(3):117-128\nhttps://doi.org/10.23838/pfm.2025.00177\npISSN: 2508-7940 · eISSN: 2508-7959\n1 / 1CROSSMARK_logo_3_Test\n2017-03-16https://crossmark-cdn.crossref.org/widget/v2.0/logos/CROSSMARK_Color_square.svg\nReceived: June 22, 2025 \nRevised: September 1, 2025\nAccepted: September 6, 2025 \n \nCorresponding author: \nShady Saud Khan\nGeneral Medicine Practice \nProgram, Batterjee Medical \nCollege, Jeddah 21442, Saudi \nArabia\nTel: +966-556550655\nE-mail: shadysaudk@gmail.com\n\n118 http://pfmjournal.org\nMachine learning in endometriosis detection\nINTRODUCTION\nEndometriosis is a chronic, inflammatory gynecological dis-\norder characterized by the presence of endometrial-like tis-\nsue outside the uterus, which leads to pain, infertility, and re-\nduced quality of life. It affects approximately 10% of repro-\nductive-aged women worldwide, with some estimates sug-\ngesting an even higher prevalence due to underdiagnosis [1]. \nThe most common symptoms include dysmenorrhea (painful \nmenstruation), chronic pelvic pain, dyspareunia (pain during \nintercourse), and subfertility, although symptom severity \ndoes not always correlate with disease stage [2]. In addition \nto physical suffering, endometriosis significantly affects men-\ntal health, work productivity, and social relationships, with \nmany patients reporting anxiety, depression, and reduced \noverall well-being [3]. The economic burden is also consider-\nable, with healthcare expenditures and productivity losses \namounting to billions of dollars annually [4]. Given its wide-\nspread impact, enhancing diagnostic and therapeutic strate-\ngies is critical for mitigating the long-term consequences of \nthis debilitating condition.\nDespite its prevalence, endometriosis remains notoriously \ndifficult to diagnose, with an average delay of 7 to 10 years \nfrom symptom onset to definitive diagnosis [5]. The gold \nstandard for diagnosis is laparoscopic surgery with histologi-\ncal confirmation, an invasive procedure that carries risks such \nas infection, bleeding, and complications related to anesthe-\nsia [6]. Noninvasive imaging techniques, including transvagi-\nnal ultrasound (TVUS) and magnetic resonance imaging \n(MRI), are frequently employed but demonstrate variable \nsensitivity and specificity depending on lesion type and oper-\nator expertise [7]. Furthermore, biomarkers such as cancer \nantigen 125 (CA-125) lack sufficient accuracy for use as stand-\nalone diagnostic tools, leading to frequent misdiagnoses with \nconditions such as irritable bowel syndrome or pelvic inflam-\nmatory disease [8]. These diagnostic limitations underscore \nthe urgent need for more reliable and less invasive methods \nto reduce delays and improve patient outcomes. The pro-\ntracted diagnostic journey for endometriosis patients high-\nlights the importance of developing early, noninvasive diag-\nnostic tools. Early detection could prevent disease progres-\nsion, reduce chronic pain, and preserve fertility by enabling \ntimely intervention [9]. Emerging research is investigating liq-\nuid biopsies, proteomics, and microRNA profiling as potential \ndiagnostic strategies, although none have yet achieved wide-\nspread clinical adoption [10]. In addition, integrating pa-\ntient-reported symptoms with imaging findings may enhance \ndiagnostic accuracy without immediate reliance on surgery \n[11]. Given the shortcomings of current methods, there is a \npressing demand for innovative approaches that combine \nclinical data with advanced analytical techniques to stream-\nline diagnosis and improve patient care.\nArtificial intelligence (AI) and machine learning (ML) are \ntransforming healthcare by enabling automated pattern rec-\nognition, predictive modeling, and decision-support systems. \nIn medical imaging, AI algorithms have demonstrated consid-\nerable success in detecting malignancies, cardiovascular ab-\nnormalities, and neurological disorders [12]. These technolo-\ngies excel in processing large datasets, identifying subtle pat-\nterns, and enhancing diagnostic precision capabilities that \ncould help address key challenges in endometriosis diagnosis \n[13]. For example, ML models trained on ultrasound or MRI \ndatasets can improve lesion detection, while natural language \nprocessing can extract diagnostic information from electronic \nhealth records (EHRs) [14]. As AI continues to evolve, its appli-\ncation in endometriosis holds promise for reducing diagnos-\ntic delays and improving classification accuracy.\nThis review aims to critically examine the current diagnos-\ntic challenges in endometriosis and explore the potential of \nAI-driven solutions to overcome these barriers. We evaluate \nthe limitations of existing diagnostic tools, analyze recent de-\nvelopments in biomarker research and imaging technologies, \nand discuss how AI can integrate multimodal data to improve \naccuracy. Furthermore, we highlight ongoing research, clini-\ncal applications, and future directions for AI in endometriosis \ncare. By synthesizing the latest evidence, this review seeks to \ninform clinicians, researchers, and policymakers about the \ntransformative potential of AI in addressing one of gynecolo-\ngy’s most persistent diagnostic challenges.\nMATERIALS AND METHODS\nElectronic databases, including PubMed/MEDLINE, Web of \nScience, and Google Scholar, were searched for case reports, \ncase series, case-control studies, cohort studies, cross-sec-\ntional studies, and review articles from each database’s incep-\ntion to May 2025. The search terms included ‘Endometriosis,’ \n‘Machine learning,’ ‘Symptom clustering,’ and ‘Artificial intelli-\ngence,’ among others. Studies were included if they investi-\ngated the application of AI or ML to the diagnosis or early de-\ntection of endometriosis, reported original data (including \ncase-control, cohort, cross-sectional studies, or randomized \nstudies) or systematic/narrative reviews directly relevant to \nML in endometriosis, used symptom-based, imaging-based, \n\n119https://doi.org/10.23838/pfm.2025.00177\nZiad Mumtaz Ramadan, et al.\nor multimodal datasets; and were published in English. Both \nprospective and retrospective studies were eligible, and no \nrestrictions were placed on patient age or geographic setting. \nThe majority of the studies included were small to medi-\num-sized, often retrospective single-center registries with less \nthan 100 up to several hundred patients. The patient popula-\ntions were enrolled predominantly from the gynecology spe-\ncialty clinics and may have had limited generalizability to \nmore prevalent community practice. In particular, most stud-\nies defined endometriosis by surgical/pathologic diagnosis, \nthe current gold standard for diagnosis, while others used \nclinical diagnosis or imaging study findings such as ultra-\nsound or MRI. This was conducted as a narrative review rather \nthan a systematic review; therefore, no formal quality assess-\nment or strict exclusion criteria were applied. Study inclusion \nwas primarily based on relevance to the research question \nand the authors’ judgment of contribution to the evolving \nfield.\nDIAGNOSTIC CHALLENGES IN  \nENDOMETRIOSIS\nHeterogeneity of symptoms\nOne of the primary obstacles in diagnosing endometriosis is \nthe wide variability in clinical presentation. While some pa-\ntients experience severe pelvic pain and infertility, others re-\nmain asymptomatic, with lesions discovered incidentally \nduring surgery for unrelated conditions [15]. Symptom pat-\nterns also differ by lesion location; for instance, deep infiltrat-\ning endometriosis (DIE) may cause bowel or urinary symp-\ntoms, whereas ovarian endometriomas are more commonly \nassociated with dysmenorrhea [16]. This heterogeneity com-\nplicates diagnosis, as clinicians must distinguish endometrio-\nsis from other conditions with overlapping symptoms, such \nas adenomyosis, interstitial cystitis, and gastrointestinal (GI) \ndisorders [17]. The absence of a definitive symptom profile \nnecessitates a multidisciplinary diagnostic approach, further \ndelaying identification and treatment.\nAverage diagnostic delay (typically 7 to 10 years)\nThe prolonged diagnostic delay in endometriosis typically \nranging from 7 to 10 years reflects systemic shortcomings in \nawareness, clinical evaluation, and diagnostic practices [18]. \nStudies show that adolescents and young women often expe-\nrience even longer delays, as their symptoms are frequently \ndismissed as ‘normal’ menstrual pain [18,19]. Contributing \nfactors include clinician bias, normalization of pelvic pain, \nand limited access to specialized care [20]. These delays per-\nmit disease progression, thereby increasing the risk of chronic \npain, infertility, and surgical complications [21]. Reducing this \ndelay requires enhanced clinician education, standardized \nsymptom-assessment tools, and improved diagnostic tech-\nnologies to support early intervention.\nLimitations of current diagnostic tools (ultrasound, \nMRI, and CA-125)\nAlthough TVUS and MRI are widely used for endometriosis \ndetection, their diagnostic accuracy depends on lesion type, \noperator expertise, and imaging protocols [22]. TVUS is effec-\ntive for detecting ovarian endometriomas but is less sensitive \nfor superficial peritoneal or deep infiltrating lesions [23]. MRI \nprovides superior soft-tissue contrast but is costly and not \nuniversally accessible [24]. Serum biomarkers such as CA-125 \nare often elevated in endometriosis but lack specificity, as \nthey may also increase in ovarian cancer, pregnancy, and oth-\ner inflammatory conditions [25]. These limitations under-\nscore the need for more reliable and cost-effective diagnostic \ntools that can detect endometriosis across its diverse mani-\nfestations without invasive procedures.\nImportance of integrating symptomatology and  \nimaging\nGiven the limitations of individual diagnostic methods, a mul-\ntimodal approach that combines symptom assessment, im-\naging, and biomarker analysis may improve diagnostic accu-\nracy. Recent studies suggest that structured symptom ques-\ntionnaires (e.g., the Endometriosis Symptom Diary) can aid in \nstratifying patients for further evaluation [26]. When paired \nwith AI-enhanced imaging analysis, these tools could facilitate \nearlier identification of high-risk individuals [27]. For instance, \nML algorithms can process ultrasound or MRI data to detect \nsubtle lesions that may be overlooked by human evaluators \n[28]. Integrating patient history with AI-driven imaging inter-\npretation has the potential to reduce reliance on laparoscopy, \nthereby enabling faster and less invasive diagnoses. Future \nresearch should prioritize validating such integrated models \nin diverse clinical settings to ensure broad applicability.\nMACHINE LEARNING IN MEDICINE:  \nA PRIMER\nML is a subfield of AI that encompasses computational meth-\nods enabling computers to learn from medical records and \nprior experiences to categorize or predict outcomes without \n\n120 http://pfmjournal.org\nMachine learning in endometriosis detection\nexplicit programming [29,30]. The three primary types of ML \nare supervised learning, unsupervised learning, and deep \nlearning. Deep learning is a more advanced approach that \nbuilds upon the first two paradigms [31,32]. To predict out-\ncomes for new, unseen data, supervised learning requires al-\ngorithms trained on labeled datasets, in which the correct \noutputs are already known. This method is frequently applied \nin diagnostic contexts for tasks such as regression and classi-\nfication, where prior patient data are used to train models to \ndifferentiate between diseased and healthy states [31,33]. \nUnsupervised learning, by contrast, does not utilize labeled \ndata but instead identifies hidden patterns or clusters. This \napproach is particularly valuable for uncovering phenotypes \nor disease subtypes that may not be immediately apparent to \nclinicians [31,34]. Deep learning, also referred to as deep neu-\nral network learning, is especially powerful for analyzing \nmedical images and signals, employing multilayered neural \nnetworks to automatically extract features and recognize \ncomplex patterns [35,36].\nML is increasingly applied in clinical settings because it can \nidentify patterns within large and complex medical datasets. \nML algorithms can assist clinicians in developing more effec-\ntive treatment strategies and making more accurate diagno-\nses by analyzing patient information. For example, an ML \nmodel can use hospital data on symptoms and treatment \noutcomes to predict which therapies are most likely to be ef-\nfective for specific patient groups [35,37]. These algorithms, \noften referred to as classifiers, can accurately distinguish be-\ntween diagnostic categories depending on the model archi-\ntecture and the quality of the data [37,38]. ML is also used for \nrisk prediction, employing algorithms such as support vector \nmachines (SVMs) and convolutional neural networks (CNNs) \nto identify high-risk patients at earlier stages [31,38]. When ap-\nplied collectively, these approaches enable clinicians to detect \ndiseases sooner, design personalized treatment plans, and \nachieve improved outcomes across diverse medical fields.\nOne area where ML has demonstrated particular value is \ngynecology, where it enhances the accuracy and efficiency of \ndisease detection, risk assessment, and clinical decision-mak-\ning, especially in cancer diagnosis and pregnancy-related \ncomplications [39,40]. In cancer care, ML and deep learning \nmodels are applied to diagnose cervical, ovarian, and endo-\nmetrial cancers through image analysis, biomarker discovery, \nand individualized risk prediction [41]. These models fre-\nquently outperform conventional diagnostic methods in sen-\nsitivity and accuracy, thereby facilitating earlier detection and \nreducing treatment delays [41-43].\nBeyond oncology, ML also plays an important role in pre-\ndicting and managing a wide range of gynecological condi-\ntions. Algorithms such as decision tree ensembles and naïve \nBayes classifiers have demonstrated high effectiveness in pre-\ndicting conditions including postpartum depression, anemia, \npreeclampsia, gestational diabetes, and miscarriage achiev-\ning accuracies of up to 86% using symptom and clinical data \n[44-46]. In maternal-fetal medicine, ML facilitates the early \nidentification of complications such as preterm birth, cesare-\nan delivery, and perinatal death, often through the integration \nof clinical metrics with imaging data [40,44]. Furthermore, AI \nand ML are transforming gynecologic imaging by automating \nimage classification, assisting radiologists in their diagnostic \ntasks, and promoting greater consistency and objectivity in \nclinical decision-making [42,43]. These developments illus-\ntrate how the fundamental principles of ML are being applied \nto integrate clinical care with computational methods, there-\nby achieving tangible improvements in women’s health.\nSYMPTOM CLUSTERING IN  \nENDOMETRIOSIS USING MACHINE \nLEARNING\nML-based symptom clustering has emerged as a viable strate-\ngy for characterizing the complex clinical manifestations of \nendometriosis. Infertility, pelvic pain, dysmenorrhea, dyspa-\nreunia, and GI symptoms are frequently included as key vari-\nables for clustering and prediction in endometriosis research \n(Table 1). While GI symptoms are often misattributed, contrib-\nuting to diagnostic delays, pelvic pain and dysmenorrhea are \nparticularly prevalent in both adolescent and adult popula-\ntions. These symptoms are central to patient-reported out-\ncomes as well as clinical data analyses [47-50].\nBuilding on this symptom-based framework, unsupervised \nML techniques such as multivariate mixture models, parti-\ntioning around medoids (PAM), and k-means clustering have \nbeen employed to identify subtypes within endometriosis \npopulations. K-means clustering, for example, has been ap-\nplied to classify women according to quality-of-life parame-\nters, revealing discrete groups with either high or low quality \nof life [51]. Similarly, analyses of adolescent EHR data using \nPAM and mixture models have identified ‘classic’ (pelvic pain, \ndysmenorrhea, chronic pain), ‘GI’ (gastrointestinal-domi-\nnant), and ‘feature-absent’ phenotypes [49]. In addition, ana-\ntomical pain sites have been mapped using Bayesian network \nanalysis, which has also been applied to identify patterns in \npain symptomatology [47]. Furthermore, mixed-membership \n\n121https://doi.org/10.23838/pfm.2025.00177\nZiad Mumtaz Ramadan, et al.\nmodels have shown promise in handling multimodal, self-\ntracked symptom data, thereby improving the clinical rele-\nvance of identified subtypes [50].\nIncreasingly, studies incorporate EHR data and patient-re-\nported outcomes obtained from large-scale surveys and \nsymptom-tracking applications to support these computa-\ntional approaches and capture the full heterogeneity of endo-\nmetriosis. These data sources enable researchers to identify \nphenotypes directly from real-world experiences through un-\nsupervised learning [50,51]. Clinical documentation has also \nbeen analyzed with EHR-based clustering, which has revealed \na spectrum of symptom profiles and their associations with \ntreatment patterns [49].\nCrucially, these clustering models not only help define en-\ndometriosis subtypes but also aid in distinguishing the condi-\ntion from other pelvic pain disorders. ML techniques have \ndemonstrated higher predictive value for diagnosing endo-\nmetriosis compared with other pain syndromes by identifying \nspecific symptom constellations, such as the co-occurrence \nof chronic pelvic pain, subfertility, and dyspareunia [47]. Fur-\nthermore, studies indicate that women with classic pheno-\ntypes typically receive more hormonal or pain-related inter-\nventions, whereas women with non-classic or GI-dominant \npresentations may receive less clinical attention [49,50]. \nThese findings underscore the clinical value of detailed symp-\ntom profiling in improving differential diagnosis and guiding \nindividualized therapeutic approaches.\nIMAGING-BASED MACHINE LEARNING \nMODELS\nTVUS and MRI are among the most widely used noninvasive \nmethods for detecting endometriosis, with each offering dis-\ntinct advantages and disadvantages. TVUS is a first-line imag-\ning modality for endometriosis because of its minimal risk, \nlow cost, and high diagnostic performance, with reported \nsensitivity ranging from 71% to 98% and specificity from 92% \nto 100%. TVUS can detect endometriotic lesions in the uterine \nlining as small as 5 mm, typically appearing as hypoechoic \nmasses, while peritoneal lesions larger than 5 mm are classi-\nfied as DIE. The ‘tenderness-guided’ technique and the ‘slid-\ning sign’ are commonly used to assess lesion depth and the \npresence of endometriosis in the pouch of Douglas (POD) [52].\nMRI is considered a second-line imaging modality following \nTVUS due to its higher cost and limited availability. However, \nMRI offers important advantages, including the ability to eval-\nuate larger pelvic structures such as the bowel, ureters, and \nextra-pelvic lesions, as well as high sensitivity for fibrin degra-\ndation products that facilitate the detection of obscured le-\nsions [52]. Although TVUS and MRI remain essential in endo-\nmetriosis diagnosis, image interpretation is challenging. In \nthis context, ML models such as CNNs and radiomics ap -\nproaches have shown potential for improving diagnostic ac-\ncuracy.\nCNNs and radiomics as tools in medical imaging\nDiagnosing endometriosis using noninvasive methods such \nas TVUS and MRI remains challenging, with inconsistencies \neven in expert interpretation. However, the emergence of \nAI-driven, imaging-based ML models such as CNNs and radio-\nmics offers the potential for more accurate and timely diag-\nnosis when trained on large datasets of annotated images \nthat capture subtle patterns of endometriosis often imper-\nceptible to the human eye [53]. CNNs are a class of deep \nlearning models capable of processing vast amounts of data; \nthey are specifically designed for tasks such as image pattern \nTable 1. Summary of symptom clustering in endometriosis using machine learning \nKey aspect Methods used              Data source                            Main findings/insights\nKey symptoms Pelvic pain, dysmenorrhea,  \ndyspareunia, infertility, GI issues\nPatient-reported, EHRs Central to clustering and prediction models [47-51]\nUnsupervised learning k-means, PAM, mixture models, \nBayesian networks\nSurveys, EHRs, self-tracking Identified classic, GI, and feature-absent phenotypes; \nrobust to data variability [47,49-51]\nPatient-reported/EHR  \n   data\nSurveys, self-tracking,  \nEHR notes\nLarge cohorts, clinical notes Enabled detection of symptom heterogeneity and \nless-recognized phenotypes [49,51]\nDistinguishing  \n   disorders\nClustering of symptom  \nconstellations\nBayesian networks,  \nclustering\nClassic symptom clusters increase endometriosis  \nrisk; models help differentiate from other pain  \ndisorders [47,49]\nGI, gastrointestinal; EHR, electronic health record; PAM, partitioning around medoids.\n\n122 http://pfmjournal.org\nMachine learning in endometriosis detection\nrecognition. In medical contexts, CNNs are frequently applied \nfor detection, classification, and segmentation. Among CNN-\nbased models, U-Net is particularly well suited for medical \nimage segmentation. This ‘U-shaped’ neural network has \nbeen extensively used in research on gynecological and onco-\nlogical conditions. Notably, a 2024 study demonstrated that \nU-Net achieved highly accurate segmentation of endometri-\notic lesions, with a Dice coefficient of 0.977 suggesting that \nCNN-based models trained on sufficiently large datasets can \nsubstantially improve the accuracy of noninvasive endome-\ntriosis diagnosis [53].\nIn addition to CNNs, another imaging-based AI technique is \nradiomics, a process that extracts and analyzes quantitative \nfeatures from medical images and transforms them into \nhigh-dimensional data such as shape, texture, and intensity. \nThe radiomics workflow typically involves several key steps: \nimage acquisition, preprocessing, segmentation, feature ex-\ntraction, and model building [54,55]. Radiomics can detect \nsubtle patterns and variations in endometriosis that are often \noverlooked during traditional interpretation, thereby provid-\ning a more comprehensive diagnostic approach [54,55]. For \nexample, in a study that employed radiomics to develop a \nmodel distinguishing ovarian endometriomas from dermoid \ncysts, the model achieved an area under the curve (AUC) of \n0.981 and an accuracy of approximately 94% [56]. A recent \nadvancement in this field is the Human–AI Collaborative Mul-\ntimodal Multirater Learning (HAICOMM) method, which inte-\ngrates radiomics with collaborative human–AI models to fur-\nther enhance diagnostic performance in endometriosis. HAI-\nCOMM demonstrated significant improvements in the classi-\nfication of POD obliteration on MRI images, outperforming in-\nterpretations made by either clinicians or AI models alone \n[57]. Together, CNNs and radiomics exemplify how AI holds \nsubstantial promise for improving endometriosis diagnosis \nthrough noninvasive methods, thereby reducing diagnostic \ndelays and improving patient outcomes.\nML models in detecting endometrioma, DIE, and  \nperitoneal lesions\nSuperficial peritoneal endometriosis\nIn one study, several ML algorithms including random forest \nand gradient boosting were trained on data from clinical histo-\nry and physical examinations in women with chronic pelvic \npain but no detectable abnormalities on imaging. The best- \nperforming models achieved a sensitivity of 79.3% and a spec-\nificity of 74.2% for predicting superficial peritoneal endome-\ntriosis (SPE), with subsequent surgical confirmation. Key pre-\ndictive features included oligomenorrhea, bladder pain syn-\ndrome, and irritable bowel syndrome [58]. Additionally, an-\nother study reported that a random forest model using a saliva \nmicroRNA signature (89 miRNAs) achieved 100% sensitivity, \nspecificity, and AUC for detecting the SPE phenotype. This \nfinding suggests a highly accurate and noninvasive diagnostic \ntool that could potentially replace invasive procedures such as \nlaparoscopy [59].\nDIE and endometrioma\nIn a study on DIE, the inclusion of 16 clinical symptoms in an \nML algorithm yielded AUC values ranging from 0.91 to 0.95 in \ntraining datasets and from 0.66 to 0.92 in test datasets. These \nmodels may serve as screening tools for DIE in both gyneco-\nlogical and general practice settings [28]. For endometriomas, \ndifferentiation from SPE and DIE relies on imaging modalities \nsuch as ultrasound and MRI, as well as distinct cytokine signa-\ntures in peritoneal fluid. Future ML models incorporating \nthese biomarkers have the potential to further improve diag-\nnostic accuracy [60].\nHuman vs. ML diagnostic performance\nAlthough ML models have shown promise in the medical im-\naging of endometriosis, a key question remains: do ML mod-\nels outperform clinicians in diagnosis? Comparative evalua-\ntion of human and ML diagnostic performance is still an \nemerging area of research. ML approaches aim to improve \naccuracy, speed, and noninvasiveness, whereas current clini-\ncal standards continue to rely heavily on clinical evaluation \nand invasive procedures such as laparoscopy.\nOn one hand, several studies support the diagnostic value \nof ML models using noninvasive methods. For instance, a \n2023 study employing a CNN-based deep learning model, Vi-\nsual Geometry Group (VGG-16), applied to MRI images report-\ned a sensitivity of 84.15% and a specificity of 83.86%, indicat-\ning that ML has the potential to reduce false negatives and di-\nagnostic delays [61]. Other studies applying ML algorithms to \ndetect ultrasonographic signs reported sensitivities and spec-\nificities between 70% and 80%. Well-known models such as \nrandom forest and extra-trees achieved an AUC of 0.76, sug-\ngesting moderate to good diagnostic performance [62]. On \nthe other hand, a systematic review comparing TVUS and MRI \nreported sensitivity and specificity values across different le-\nsion locations (Table 2). TVUS demonstrated similar sensitivi-\nty to MRI but slightly higher specificity. These findings suggest \nthat both methods are reliable, though performance varies \ndepending on lesion location [22].\n\n123https://doi.org/10.23838/pfm.2025.00177\nZiad Mumtaz Ramadan, et al.\nLimitations of imaging-based ML\nAlthough ML techniques have shown promising results in di-\nagnostic accuracy, traditional methods remain important in \nclinical practice. The current standard for diagnosing endo-\nmetriosis involves clinical evaluation, imaging modalities \nsuch as MRI and ultrasound, and confirmatory laparoscopy. \nThis reliance often results in diagnostic delays due to the \ncomplex and heterogeneous nature of symptoms. A major \nlimitation of ML is its tendency to overfit, particularly when \ntrained on small datasets. Consequently, researchers empha-\nsize the need for large, well-curated datasets to improve train-\ning and to account for the variability of symptoms experi-\nenced throughout a patient’s lifetime. Addressing this chal-\nlenge requires a comprehensive strategy that integrates ML \nwith imaging and additional diagnostic approaches, such as \nbiomarker testing, to more effectively detect endometriosis \n[63]. Despite encouraging findings, the field still requires sys-\ntematic protocols for image preprocessing and feature ex-\ntraction across imaging modalities, as well as clinical valida-\ntion through large-scale cohort studies to ensure reliability \nand generalizability. Collectively, current evidence suggests \nthat human interpretation remains a necessary component \nof endometriosis diagnosis, despite the limitations faced by \nclinicians. Only once modern AI technologies are consistently \nvalidated and shown to outperform conventional methods \ncan they be considered a standalone alternative.\nINTEGRATION OF SYMPTOM  \nCLUSTERING WITH IMAGING DATA\nAlthough ML models such as CNNs and radiomics have \nshown substantial potential in detecting endometriotic le-\nsions, their accuracy is limited when the heterogeneous clini-\ncal presentations of patients are not considered. To address \nthis limitation, recent studies have focused on integrating \nsymptom clustering with advanced ML imaging data. By joint-\nly evaluating patient-reported symptoms and radiological \nfindings, ML models can provide a more holistic and individu-\nalized approach to diagnosing endometriosis. This section \nexplores ML methodologies that emphasize such integrative \nstrategies in endometriosis research [64]. Multimodal ML re-\nfers to the use of algorithms that analyze multiple types of \ndata. In the context of endometriosis, this involves combining \nclinical symptoms with imaging data to improve diagnostic \noutcomes. The application of multimodal ML is particularly \nrelevant for complex and heterogeneous diseases such as en-\ndometriosis. Recent advances highlight the importance of \ncombining clinical, symptom, and imaging data using multi-\nmodal ML to achieve more accurate and comprehensive diag-\nnostic solutions. For instance, in one study that developed an \nML model incorporating both imaging data and patient-re-\nported symptoms, the results demonstrated high sensitivity \nand specificity in diagnosing endometriosis [65].\nFusion strategies, benefits, and challenges in  \nmultimodal ML\nA systematic review of fusion techniques found that early fu-\nsion in which raw data from different modalities are com-\nbined before training the model is the most commonly used \nstrategy compared with other fusion methods. This approach \nbenefits the model by enabling it to learn cross-modal repre-\nsentations from the outset [66]. Both conventional ML models \n(e.g., decision trees and SVMs) and deep learning models \n(e.g., CNNs and transformers) can be integrated into multi-\nmodal systems. Additionally, Bayesian networks have been \napplied to cluster pain symptoms and relate them to surgical \ndiagnoses, thereby improving identification of symptom pat-\nterns in endometriosis [67]. Ensemble and decision tree \nmethods are also frequently employed to combine clinical \nand imaging data, making them useful in the broader context \nof leveraging multimodal ML for disease diagnosis [47].\nThe benefits of multimodal ML include improved diagnostic \naccuracy through the integration of symptom patterns, imag-\ning features (e.g., MRI), and clinical evaluations, which helps \novercome the limitations of standalone clinician assessments \nor ML models [16]. Multimodal ML can also facilitate the iden-\ntification of distinct subtypes of endometriosis, thereby sup-\nporting earlier and more personalized treatment. Further-\nmore, ML models that integrate imaging with clinical and \nsymptom data can achieve more accurate mapping of lesion \nlocation and depth, which is critical for surgical planning [51]. \nNevertheless, several challenges remain. Variability in symp-\ntom reporting and imaging protocols may complicate the data \nprocessed by ML models. Many studies are also constrained \nby small or single-center cohorts, limiting generalizability and \nstatistical significance. Finally, the complexity of advanced ML \nTable 2. Comparison between sensitivity and specificity of magnetic \nresonance imaging in various locations\nLocation Sensitivity (%) Specificity (%)\nRecto-sigmoid 85 95 \nRecto-vaginal septum 66 97 \nUterus 70 93 \n\n124 http://pfmjournal.org\nMachine learning in endometriosis detection\nmodels such as ensembles can make them difficult for clini-\ncians to interpret and trust in decision-making [68].\nThe integration of multimodal ML has the potential to ad-\nvance current diagnostic guidelines for endometriosis, but \naddressing existing challenges and utilizing large datasets will \nbe essential for widespread clinical adoption.\nLIMITATIONS AND CURRENT GAPS IN \nTHE LITERATURE\nAlthough advances in technology applied to AI- and ML-based \napproaches can significantly reduce diagnostic delay and in-\ncrease sensitivity of noninvasive endometriosis diagnosis, \ndownstream clinical impact is complex. Treatment choice \ncontinues to remain symptomatic and empirical in practice \ntoday with excessive use of symptom suppression by hor-\nmones, analgesics, or surgery, rather than disease-pheno-\ntype-guided tailored approaches [2,9,21]. As has been em-\nphasized by recent criticism, endometriosis continues to be \nmanaged largely as chronic pain syndrome, symptom-load \ndriven treatment rather than overt diagnostic subtyping [16]. \nThus, even if ML models enhance the likelihood of earlier or \nmore accurate detection, their short-term impact on thera-\npeutic gain may be blunted unless diagnoses progress are in-\ntegrated into evidence-based treatment pathways. Follow-up \nanalyses need to move beyond initial trend toward diagnosis \nitself to ascertain if early diagnosis with AI is being translated \ninto better clinical results, such as reduced pain, preservation \nof fertility, better quality of life, or more personalized opera-\ntive planning. Without this translation step, better diagnostic \nequipment is technology advancement with no-holds-barred \ntherapeutic correlations. Despite the promising advances in \napplying ML to the early diagnosis of endometriosis, several \nlimitations and gaps remain in the current literature. One of \nthe most significant limitations is the lack of large, labeled \ndatasets that combine clinical symptoms with imaging data, \nwhich hinders the development and validation of robust pre-\ndictive models. Most recent studies have focused exclusively \non either symptom clustering or imaging interpretation, rath-\ner than leveraging the synergistic benefits of multimodal data \nintegration. For example, Wang et al. [57] introduced the  \nHAICOMM model to address this gap by combining T1- and \nT2-weighted MRI images with domain-annotator labels, but \nsuch efforts remain scarce in the literature. Generalizability of \nexisting models is further limited by reliance on single-institu-\ntion retrospective datasets, which may not adequately repre-\nsent heterogeneous patient populations. Maleki et al. [69] \nemphasized that methodological shortcomings such as data \nleakage and batch effects can lead to overestimation of mod-\nel performance and reduce applicability across different clini-\ncal settings. In addition, the lack of standardized evaluation \nmetrics and benchmarking protocols hinders meaningful \ncomparison of model performance across studies [69]. Sala-\nhuddin et al. [70] advocated for uniform evaluation frame-\nworks to ensure both interpretability and reliability of deep \nlearning models in medical imaging. Interpretability itself re-\nmains a major challenge for complex models such as deep \nlearning networks, with limited efforts directed toward incor-\nporating explainability tools that could foster clinician trust \nand usability [70]. Solutions such as InterpretML (Microsoft \nResearch) have been proposed, but their application to endo-\nmetriosis research is only beginning to emerge. Addressing \nthese challenges will be critical to transitioning ML-based \ntools from experimental applications to routine clinical prac-\ntice in endometriosis diagnosis [71]. \nCONCLUSION\nML represents a transformative approach to advancing early \ndiagnosis of endometriosis by integrating symptom cluster-\ning with imaging data. Current models show strong potential; \nhowever, their translation into clinical practice is significantly \nlimited by challenges in data availability, interpretability, and \nstandardization. To encourage widespread clinical adoption \nat an accelerated pace, studies must continue to tackle; creat-\ning multicenter, multimodal data consortia with wide and \nheterogeneous datasets, standardized benchmarking plat-\nforms and assessment measures to make studies compara-\nble, incorporating explainability tools to establish clinician \ntrust and acceptance into routine operating practice, con-\nducting prospective external validation studies to determine \nreal-world generalizability and conducting cost-effectiveness \nanalyses to inform policy planning and the use of healthcare \nresources. By meeting these priorities, ML-based diagnostic \ntools can move on from proof-of-concept to clinically viable \nsolutions and ultimately reduce diagnostic delays and im-\nproved outcomes for endometriosis in women.\nCONFLICTS OF INTEREST\nNo potential conflict of interest relevant to this article was re-\nported.\n\n125https://doi.org/10.23838/pfm.2025.00177\nZiad Mumtaz Ramadan, et al.\nACKNOWLEDGMENTS\nThe authors would like to thank Dr Noha Farag for her valu-\nable time and contribution.\nORCID\nZiad Mumtaz Ramadan \n     https://orcid.org/0009-0001-7436-2958\nSaaid Mounzer Mouazen \n     https://orcid.org/0009-0001-1013-9343\nShady Saud Khan    https://orcid.org/0009-0005-7684-5426\nSariya Khan     https://orcid.org/0009-0003-9809-872X\nNoha S. Farag     https://orcid.org/0009-0001-9549-6832\nAUTHOR CONTRIBUTIONS\nConception or design: ZMR, SMM, SSK, SK, NSF.\nAcquisition, analysis, or interpretation of data: ZMR, SMM, \nSSK, SK, NSF.\nDrafting the work or revising: ZMR, SMM, SSK, SK, NSF.\nFinal approval of the manuscript: ZMR, SMM, SSK, SK, NSF.\nREFERENCES \n1. As-Sanie S, Mackenzie SC, Morrison L, Schrepf A, Zonder-\nvan KT, Horne AW, et al. Endometriosis: a review. JAMA \n2025;334:64-78.\n2. Taylor HS, Kotlyar AM, Flores VA. Endometriosis is a chron-\nic systemic disease: clinical challenges and novel innova-\ntions. Lancet 2021;397:839-52.\n3. Brauner EV, Koch T, Juul A, Doherty DA, Hart R, Hickey M. \nPrenatal exposure to maternal stressful life events and \nearlier age at menarche: the Raine Study. Hum Reprod \n2021;36:1959-69.\n4. Simoens S, Dunselman G, Dirksen C, Hummelshoj L, Bokor \nA, Brandes I, et al. The burden of endometriosis: costs and \nquality of life of women with endometriosis and treated in \nreferral centres. Hum Reprod 2012;27:1292-9.\n5. Agarwal SK, Chapron C, Giudice LC, Laufer MR, Leyland N, \nMissmer SA, et al. Clinical diagnosis of endometriosis: a \ncall to action. Am J Obstet Gynecol 2019;220:354.\n6. Johnson NP, Hummelshoj L, Adamson GD, Keckstein J, \nTaylor HS, Abrao MS, et al. World Endometriosis Society \nconsensus on the classification of endometriosis. Hum \nReprod 2017;32:315-24.\n7. Guerriero S, Condous G, van den Bosch T, Valentin L, Le-\none FP, Van Schoubroeck D, et al. Systematic approach to \nsonographic evaluation of the pelvis in women with sus-\npected endometriosis, including terms, definitions and \nmeasurements: a consensus opinion from the Interna-\ntional Deep Endometriosis Analysis (IDEA) group. Ultra-\nsound Obstet Gynecol 2016;48:318-32.\n8. Nisenblat V, Prentice L, Bossuyt PM, Farquhar C, Hull ML, \nJohnson N. Combination of the non-invasive tests for the \ndiagnosis of endometriosis. Cochrane Database Syst Rev \n2016;7:CD012281.\n9. Vercellini P, Somigliana E, Vigano P, Abbiati A, Daguati R, \nCrosignani PG. Endometriosis: current and future medical \ntherapies. Best Pract Res Clin Obstet Gynaecol 2008;22: \n275-306.\n10. Suryawanshi S, Vlad AM, Lin HM, Mantia-Smaldone G, Las-\nkey R, Lee M, et al. Plasma microRNAs as novel biomark-\ners for endometriosis and endometriosis-associated ovar-\nian cancer. Clin Cancer Res 2013;19:1213-24.\n11. Leonardi M, Uzuner C, Mestdagh W, Lu C, Guerriero S, Za-\njicek M, et al. Diagnostic accuracy of transvaginal ultra-\nsound for detection of endometriosis using International \nDeep Endometriosis Analysis (IDEA) approach: prospec-\ntive international pilot study. Ultrasound Obstet Gynecol \n2022;60:404-13.\n12. Esteva A, Robicquet A, Ramsundar B, Kuleshov V, DePristo \nM, Chou K, et al. A guide to deep learning in healthcare. \nNat Med 2019;25:24-9.\n13. Topol EJ. High-performance medicine: the convergence \nof human and artificial intelligence. Nat Med 2019;25:44-\n56.\n14. Bini SA. Artificial intelligence, machine learning, deep \nlearning, and cognitive computing: what do these terms \nmean and how will they impact health care? J Arthroplas-\nty 2018;33:2358-61.\n15. Coccia ME, Nardone L, Rizzello F. Endometriosis and infer-\ntility: a long-life approach to preserve reproductive integ-\nrity. Int J Environ Res Public Health 2022;19:6162.\n16. Chapron C, Marcellin L, Borghese B, Santulli P. Rethinking \nmechanisms, diagnosis and management of endometri-\nosis. Nat Rev Endocrinol 2019;15:666-82.\n17. Stratton P, Berkley KJ. Chronic pelvic pain and endometri-\nosis: translational evidence of the relationship and impli-\ncations. Hum Reprod Update 2011;17:327-46.\n18. Surrey E, Soliman AM, Trenz H, Blauer-Peterson C, Sluis A. \nImpact of endometriosis diagnostic delays on healthcare \nresource utilization and costs. Adv Ther 2020;37:1087-99.\n19. Ballard K, Lowton K, Wright J. What’s the delay?: a quali-\n\n126 http://pfmjournal.org\nMachine learning in endometriosis detection\ntative study of women’s experiences of reaching a diagno-\nsis of endometriosis. Fertil Steril 2006;86:1296-301.\n20. Soliman AM, Fuldeore M, Snabes MC. Factors associated \nwith time to endometriosis diagnosis in the United States. \nJ Womens Health (Larchmt) 2017;26:788-97.\n21. Vercellini P, Vigano P, Somigliana E, Fedele L. Endometrio-\nsis: pathogenesis and treatment. Nat Rev Endocrinol 2014; \n10:261-75.\n22. Guerriero S, Saba L, Pascual MA, Ajossa S, Rodriguez I, \nMais V, et al. Transvaginal ultrasound vs magnetic reso-\nnance imaging for diagnosing deep infiltrating endome-\ntriosis: systematic review and meta-analysis. Ultrasound \nObstet Gynecol 2018;51:586-95.\n23. Bazot M, Bharwani N, Huchon C, Kinkel K, Cunha TM, \nGuerra A, et al. European Society of Urogenital Radiology \n(ESUR) guidelines: MR imaging of pelvic endometriosis. \nEur Radiol 2017;27:2765-75.\n24. Nisenblat V, Bossuyt PM, Farquhar C, Johnson N, Hull ML. \nImaging modalities for the non-invasive diagnosis of endo-\nmetriosis. Cochrane Database Syst Rev 2016;2:CD009591.\n25. Maksym RB, Hoffmann-Mlodzianowska M, Skibinska M, \nRabijewski M, Mackiewicz A, Kieda C. Immunology and \nimmunotherapy of endometriosis. J Clin Med 2021;10: \n5879.\n26. Nnoaham KE, Hummelshoj L, Kennedy SH, Jenkinson C, \nZondervan KT. Developing symptom-based predictive \nmodels of endometriosis as a clinical screening tool: re-\nsults from a multicenter study. Fertil Steril 2012;98:692-\n701.\n27. Cetera GE, Tozzi AE, Chiappa V, Castiglioni I, Merli CE, Ver-\ncellini P. Artificial intelligence in the management of wom-\nen with endometriosis and adenomyosis: can machines \never be worse than humans? J Clin Med 2024;13:2950.\n28. Bendifallah S, Puchar A, Suisse S, Delbos L, Poilblanc M, \nDescamps P, et al. Machine learning algorithms as new \nscreening approach for patients with endometriosis. Sci \nRep 2022;12:639.\n29. Hamet P, Tremblay J. Artificial intelligence in medicine. \nMetabolism 2017;69S:S36-40.\n30. Waring J, Lindvall C, Umeton R. Automated machine learn-\ning: review of the state-of-the-art and opportunities for \nhealthcare. Artif Intell Med 2020;104:101822.\n31. Sidey-Gibbons JA, Sidey-Gibbons CJ. Machine learning in \nmedicine: a practical introduction. BMC Med Res Meth-\nodol 2019;19:64.\n32. Garg A, Mago V. Role of machine learning in medical re-\nsearch: a survey. Comput Sci Rev 2021;40:100370.\n33. Dallora AL, Eivazzadeh S, Mendes E, Berglund J, Ander-\nberg P. Machine learning and microsimulation techniques \non the prognosis of dementia: a systematic literature re-\nview. PLoS One 2017;12:e0179804.\n34. Shameer K, Johnson KW, Glicksberg BS, Dudley JT, Sen-\ngupta PP. Machine learning in cardiovascular medicine: \nare we there yet? Heart 2018;104:1156-64.\n35. Gopinath N, Suresh Anand M, Ishwarya MV. Machine learn-\ning for medical images. In: Ganesh Babu TR, Saravanaku-\nmar U, Pattanaik B, editors. Computational imaging and \nanalytics in biomedical engineering. Apple Academic \nPress; 2024. p. 95-104. \n36. Kim H. Artificial intelligence for 6G. Springer International \nPublishing; 2022. \n37. Karako K, Tang W. Applications of and issues with machine \nlearning in medicine: bridging the gap with explainable AI. \nBiosci Trends 2025;18:497-504.\n38. Kononenko I. Machine learning for medical diagnosis: his-\ntory, state of the art and perspective. Artif Intell Med 2001; \n23:89-109.\n39. Fiste O, Liontos M, Zagouri F, Stamatakos G, Dimopoulos \nMA. Machine learning applications in gynecological can-\ncer: a critical review. Crit Rev Oncol Hematol 2022;179: \n103808.\n40. Khan I, Khare BK. Exploring the potential of machine learn-\ning in gynecological care: a review. Arch Gynecol Obstet \n2024;309:2347-65.\n41. Idlahcen F, Idri A, Goceri E. Exploring data mining and ma-\nchine learning in gynecologic oncology. Artif Intell Rev \n2024;57:20.\n42. Taddese AA, Tilahun BC, Awoke T, Atnafu A, Mamuye A, \nMengiste SA. Deep-learning models for image-based gy-\nnecological cancer diagnosis: a systematic review and \nmeta-analysis. Front Oncol 2024;13:1216326.\n43. Shrestha P, Poudyal B, Yadollahi S, Wright DE, Gregory AV, \nWarner JD, et al. A systematic review on the use of artifi-\ncial intelligence in gynecologic imaging: background, state \nof the art, and future directions. Gynecol Oncol 2022;166: \n596-605.\n44. Mennickent D, Rodriguez A, Opazo MC, Riedel CA, Castro E, \nEriz-Salinas A, et al. Machine learning applied in maternal \nand fetal health: a narrative review focused on pregnancy \ndiseases and complications. Front Endocrinol (Lausanne) \n2023;14:1130139.\n45. Cellini P, Pigoni A, Delvecchio G, Moltrasio C, Brambilla P. \nMachine learning in the prediction of postpartum depres-\nsion: a review. J Affect Disord. 2022;309:350-7.\n\n127https://doi.org/10.23838/pfm.2025.00177\nZiad Mumtaz Ramadan, et al.\n46. Islam A, Parvin R, Sultana T. Gynecological disease predic-\ntion by machine learning. World J Adv Res Rev 2024;23: \n2107-12.\n47. Kiser AC, Schliep KC, Hernandez EJ, Peterson CM, Yandell \nM, Eilbeck K. An artificial intelligence approach for investi-\ngating multifactorial pain-related features of endometrio-\nsis. PLoS One 2024;19:e0297998.\n48. Goldstein A, Cohen S. Self-report symptom-based endo-\nmetriosis prediction using machine learning. Sci Rep \n2023;13:5499.\n49. Cohen RM, Leventhal E, Nukavarapu N, Lazarov V, Hanif S, \nElovitz MA, et al. Phenotyping adolescent endometriosis: \ncharacterizing symptom heterogeneity through note-and \npatient-level clustering. medRxiv [Preprint] 2025 Feb 12. \nhttps://doi.org/10.1101/2025.02.10.25321215\n50. Urteaga I, McKillop M, Elhadad N. Learning endometriosis \nphenotypes from patient-generated data. NP J Digit Med \n2020;3:88.\n51. Vallee A, Arutkin M, Ceccaldi PF, Ayoubi JM. Quality of life \nidentification by unsupervised cluster analysis: a new ap-\nproach to modelling the burden of endometriosis. PLoS \nOne 2025;20:e0317178.\n52. Harshe G. Magnetic resonance imaging in evaluation of \nendometriosis: case-based pictorial essay. Med J Armed \nForces India 2024;80:25-32.  \n53. Ithani QT, Mokri SS, Zulkarnain N, Ahmad MFB. Deep learn-\ning model for endometrium segmentation in transvaginal \nultrasound (TVUS) images. Adv Eng Res 2024;245:66-80. \n54. Zhang W, Guo Y, Jin Q. Radiomics and its feature selec-\ntion: a review. Symmetry 2023;15:1834.\n55. Ibrahim A, Primakov S, Beuque M, Woodruff HC, Halilaj I, \nWu G, et al. Radiomics for precision medicine: current \nchallenges, future prospects, and the proposal of a new \nframework. Methods 2021;188:20-9.\n56. Liu L, Cai W, Zhou C, Tian H, Wu B, Zhang J, et al. Ultrasound \nradiomics-based artificial intelligence model to assist  \nin the differential diagnosis of ovarian endometrioma \nand ovarian dermoid cyst. Front Med (Lausanne) 2024;11: \n1362588.\n57. Wang H, Butler D, Zhang Y, Avery J, Knox S, Ma C, et al. Hu-\nman-AI collaborative multi-modal multi-rater learning for \nendometriosis diagnosis. Phys Med Biol 2024;70:015008.\n58. Santos LL, de Azevedo MC, Shimamura LK, Nogueira AA, \nCandido-dos-Reis F J, Schor E, et al. Machine learning as a \nclinical decision support tool for diagnosing superficial \nperitoneal endometriosis in women with dysmenorrhea \nand acyclic pelvic pain. Med Res Arch 2024;12:1-24.\n59. Bendifallah S, Dabi Y, Suisse S, Ilic J, Delbos L, Poilblanc M, \net al. Saliva-based microRNA diagnostic signature for the \nsuperficial peritoneal endometriosis phenotype. Eur J \nObstet Gynecol Reprod Biol 2024;297:187-96.\n60. Zhou J, Chern BS, Barton-Smith P, Phoon JW, Tan TY, Viar-\ndot-Foucault V, et al. Peritoneal fluid cytokines reveal new \ninsights of endometriosis subphenotypes. Int J Mol Sci \n2020;21:3515.\n61. Ribeiro WR, da Silva IFS, Diniz JOB, Silva AC, de Paiva AC, \nBrandao Salomao ACC, et al. Deep learning-based com-\nputational approach for the diagnosis of deep endometri-\nosis using magnetic resonance imaging. In: Proceedings \nof the 23rd Brazilian Symposium on Computing Applied \nto Health (SBCAS 2023); 2023 Jun 27–30; São Paulo, Brazil.\n62. Nouri B, Hashemi SH, J Ghadimi D, Roshandel S, Akhlagh-\ndoust M. Machine learning-based detection of endometri-\nosis: a retrospective study in a population of Iranian fe-\nmale patients. Int J Fertil Steril 2024;18:362-6.\n63. Shrestha P, Shrestha B, Sherestha J, Chen J. Current sta-\ntus and future potential of machine learning in diagnostic \nimaging of endometriosis: a literature review. JNMA J Ne-\npal Med Assoc 2025;63:205-11.\n64. Sivajohan B, Elgendi M, Menon C, Allaire C, Yong P, Bedai-\nwy MA. Clinical use of artificial intelligence in endometrio-\nsis: a scoping review. NP J Digit Med 2022;5:109.\n65. Enamorado-Diaz E, Morales-Trujillo L, Garcia-Garcia JA, \nMarcos AT, Navarro-Pando J, Escalona-Cuaresma MJ. A \nnovel machine learning-based proposal for early predic-\ntion of endometriosis disease. Expert Syst Appl 2025;271: \n126621.\n66. Mohsen F, Ali H, El Hajj N, Shah Z. Artificial intelligence- \nbased methods for fusion of electronic health records and \nimaging data. Sci Rep 2022;12:17981.\n67. Ben-Miled Z, Shebesh JA, Su J, Dexter PR, Grout RW, \nBoustani MA. Multi-modal fusion of routine care electron-\nic health records (EHR): a scoping review. Information \n(Basel) 2025;16:16010054.\n68. Quesada J, Harma K, Reid S, Rao T, Lo G, Yang N, et al. En-\ndometriosis: a multimodal imaging review. Eur J Radiol \n2023;158:110610.\n69. Maleki F, Ovens K, Gupta R, Reinhold C, Spatz A, Forghani \nR. Generalizability of machine learning models: quantita-\ntive evaluation of three methodological pitfalls. Radiol Ar-\ntif Intell 2022;5:e220028.\n70. Salahuddin Z, Woodruff HC, Chatterjee A, Lambin P. Trans-\nparency of deep neural networks for medical image anal-\nysis: a review of interpretability methods. Comput Biol \n\n128 http://pfmjournal.org\nMachine learning in endometriosis detection\nMed 2022;140:105111.\n71. Jenkins S, Nori H, Koch P, Caruana R. interpret 0.1.34: fit \ninterpretable machine learning models [Internet]. The \nComprehensive R Archive Network; 2019 [cited 2025 Sep \n1]. Available from: https://cran.r-project.org/src/contrib/\nArchive/interpret","source_license":"CC0","license_restricted":false}