Investigating the role of artificial intelligence in the diagnosis and prediction of endometriosis using ultrasound images: a systematic review

This study was conducted as a systematic review with a comprehensive literature search performed in reputable databases including PubMed, Scopus, Web of Science, and the Google Scholar search engine, using the keywords “artificial intelligence,” “prediction,” “diagnosis,” “endometriosis,” and “sonographic images,” as well as their associated MeSH terms. The search was finalized on April 30, 2025, and included all studies published up to that date. (Table  1 ) Table 1 Search strategy for each database ‎ # PubMed database search approach Results 1 (((((((“Artificial Intelligence“[Mesh]) OR “Machine Learning“[Mesh]) OR “Deep Learning“[Mesh]) OR (((“Artificial Intelligence“[Title/Abstract]) OR (“Machine Learning“[Title/Abstract])) OR (“Deep Learning“[Title/Abstract]))) AND ((“Diagnosis“[Mesh]) OR (“Diagnosis“[Title/Abstract]))) OR ((“Prognosis“[Mesh]) OR (“Prognosis“[Title/Abstract]))) AND ((“Endometriosis“[Mesh]) OR (“Endometriosis“[Title/Abstract]))) AND ((“Ultrasonography“[Mesh]) OR (“Ultrasonography“[Title/Abstract])) 281 # Scopus database search approach Results 1 ((TITLE-ABS-KEY (“Artificial Intelligence”)) OR (TITLE-ABS-KEY (“Machine learning”)) OR (TITLE-ABS-KEY (“deep learning”))) AND (TITLE-ABS-KEY (“Diagnosis”)) AND ((TITLE-ABS-KEY (“Prediction”)) OR (TITLE-ABS-KEY (“Prognosis”))) AND (TITLE-ABS-KEY (“Endometriosis”)) AND (TITLE-ABS-KEY (“Ultrasonography”)) 2 # Google Scholar database search approach Results 1 “Artificial Intelligence” OR “Machine Learning” OR “Deep Learning” AND “Diagnosis” OR “Prediction” AND “Endometriosis” AND “Ultrasound Images” 559 # Web of Science database search approach Results 1 1. TS = (“Artificial intelligence” OR “machine learning” OR “deep learning”) 2. TS = (“prognosis” OR “prediction”) 3. TS = (“diagnosis”) 4. TS = (“endometriosis”) 5. TS = (“ultrasonography”) 6. #1 AND #2 AND #3 AND #4 AND #5 1 Search strategy for each database ‎ 1. TS = (“Artificial intelligence” OR “machine learning” OR “deep learning”) 2. TS = (“prognosis” OR “prediction”) 3. TS = (“diagnosis”) 4. TS = (“endometriosis”) 5. TS = (“ultrasonography”) 6. #1 AND #2 AND #3 AND #4 AND #5 In this systematic review, only studies meeting predefined inclusion criteria were considered for full-text assessment. The inclusion criteria were as follows: original research articles published in peer-reviewed journals, available in English, and specifically focusing on the use of ultrasound imaging for the diagnosis or prediction of endometriosis with the application of artificial intelligence algorithms. Only studies reporting sufficient diagnostic performance data, such as sensitivity, specificity, or accuracy, were included to ensure scientific rigor and validity. Conversely, studies were excluded if they met any of the following criteria: non-English full texts, non-journal publications (including conference abstracts, books, review articles, letters, and correspondence to the editor), lack of alignment between the study objectives and the content of the title, abstract, or full text, use of diagnostic or predictive modalities other than ultrasound imaging, absence of artificial intelligence algorithms in the study, or lack of free full-text access. This systematic review was conducted and reported in strict accordance with the PRISMA 2020 guidelines [ 17 ]. The PRISMA checklist was followed to ensure transparency and reproducibility at each stage. The review process included four main phases: Identification, Screening, Eligibility, and Inclusion, as illustrated in the PRISMA flow diagram (Fig. 1 ). Fig. 1 Diagram of article selection based on flowchart (PRISMA) Diagram of article selection based on flowchart (PRISMA) The methodological quality and risk of bias of the included studies were assessed using the Quality Assessment of Diagnostic Accuracy Studies 2 (QUADAS-2) tool. This framework is specifically designed to evaluate diagnostic accuracy studies and is considered the gold standard for assessing bias in such research. The QUADAS-2 tool evaluates four key domains: patient selection, index test, reference standard, and flow and timing. Each domain was independently assessed by two reviewers and categorized as having a low, high, or unclear risk of bias. Any disagreements between reviewers were resolved through discussion or, when necessary, by consultation with a third reviewer. Due to considerable heterogeneity among the included studies in terms of methodology, AI algorithms, ultrasound imaging techniques, and diagnostic performance metrics, conducting a meta-analysis was not feasible. Therefore, a qualitative synthesis approach was applied. The results of the included studies were systematically compared based on AI model type, diagnostic purpose, and reported diagnostic outcomes such as accuracy, sensitivity, specificity, and error rate. Key findings were narratively summarized to identify common trends, strengths, and limitations across the studies. In this systematic review, duplicate records were initially removed using EndNote software version 21x. Subsequently, the titles and abstracts of the remaining articles were independently screened by two researchers according to predefined eligibility criteria. Full-text articles that met the inclusion criteria were then independently assessed by the same researchers. ata from the included studies were extracted using a structured checklist capturing key study characteristics such as the study title, publication year, first author, country, dataset size, study objectives, AI algorithms or techniques employed, and model characteristics (Table  2 ). Table 2 Characteristics of selected studies Source (year, Country) Aim of Study Dataset Size Scope (Diagnosis/ Prediction/ Treatment) Algorithms and/or Techniques Performance Nouri [ 18 ] Iran 2024 In this study, we aimed to compare the diagnostic accuracy of different ML algorithms for endometriosis detection. 505 patients diagnosis SVM, Random Forest, Extra-Trees, and Gradient Boosting, Nearest Neighbor, Logistic Regression, AdaBoost Algorithms Sensitivity (%) Specificity (%) PPV (%) NPV (%) AUC Nearest Neighbors 62.2 81.1 57.5 83.9 0.75 Logistic Regression 75.7 72.2 52.8 87.8 0.75 RBF SVM 70.3 71.1 50 85.3 0.76 Random Forest 67.6 77.8 55.6 85.4 0.76 Extra-Trees 70.3 80 59.1 86.7 0.76 AdaBoost 59.5 83.3 59.5 83.3 0.71 Gradient Boosting 73 73.3 52.9 86.8 0.76 Podda [ 19 ] Italy 2024 our study aimed to investigate a useful solution for the detection and segmentation of lesions in already established or strongly suspected cases of endometriosis 53 images diagnosis CNN Method Jac Dice dAcc MSUneta 0.685 0.796 0.866 DeepLabv3?b 0.5570.683 0.680 Proposed multi-scale ensemble 0.712 0.818 0.906 Guerriero [ 20 ] Italy, Spain 2021 The aim of this study was to compare the accuracy of seven classical Machine Learning (ML) models trained with ultrasound (US) soft markers to raise suspicion of endometriotic bowel involvement. 333 patients diagnosis k-NN, Naive Bayes, NNET-neuralnet, SVM, Decision Tree, Random Forest, Logistic Regression AlgorithmAccuracy Sensitivity SpecificityPPV NPV AUC k-NN0.69 0.66 0.71 0.48 0.83 75.48 Naive Bayes 0.75 0.72 0.77 0.56 0.87 80.99 NNET-neuralnet 0.73 0.72 0.73 0.52 0.86 81.73 SVM0.75 0.84 0.71 0.54 0.92 78.49 Decision Tree 0.74 0.66 0.77 0.54 0.85 77.00 Random Forest 0.70 0.66 0.72 0.49 0.84 77.18 Logistic Regression 0.73 0.72 0.73 0.52 0.86 81.37 Miao [ 21 ] China 2024 The objective of this study was to develop a deep learning model, using the ConvNeXt algorithm, that can effectively differentiate between ovarian endometriosis cysts (OEC) and benign mucinous cystadenomas (MC) by analyzing ultrasound images. 786 ultrasound images from 184 patients diagnosis deep learning model DL AUC Specificity Sensitivity PPV NPV ConvNeXt0.9090% 90% 96% 77% Cao [ 22 ] China 2025 This study aimed to explore potential factors for endometriosis severity and to develop a classification model to assess the accuracy of predicting the risk of severe endometriosis. 308 patients predicting ML algorithms, including logistic regression (LR), recursive partitioning and regression trees (rpart), random forest (RF), extreme gradient boosting (XGBoost), support vector machine (SVM), k-nearest neighbors (KNN), and neural network (NNET) set model accuracy roc_auc train logistic 0.734 0.815 train rpart 0.734 0.826 train random forest0.749 0.840 train xgboost 0.722 0.804 train svm 0.5870.813 train knn 0.710 0.779 train nnet 0.720 0.790 test logistic 0.692 0.767 test rpart 0.6030.699 test random forest 0.667 0.744 test xgboost 0.641 0.714 test svm 0.590 0.778 test knn 0.679 0.737 test nnet 0.679 0.731 Characteristics of selected studies Nouri [ 18 ] Iran 2024 SVM, Random Forest, Extra-Trees, and Gradient Boosting, Nearest Neighbor, Logistic Regression, AdaBoost Algorithms Sensitivity (%) Specificity (%) PPV (%) NPV (%) AUC Nearest Neighbors 62.2 81.1 57.5 83.9 0.75 Logistic Regression 75.7 72.2 52.8 87.8 0.75 RBF SVM 70.3 71.1 50 85.3 0.76 Random Forest 67.6 77.8 55.6 85.4 0.76 Extra-Trees 70.3 80 59.1 86.7 0.76 AdaBoost 59.5 83.3 59.5 83.3 0.71 Gradient Boosting 73 73.3 52.9 86.8 0.76 Podda [ 19 ] Italy 2024 Method Jac Dice dAcc MSUneta 0.685 0.796 0.866 DeepLabv3?b 0.5570.683 0.680 Proposed multi-scale ensemble 0.712 0.818 0.906 Guerriero [ 20 ] Italy, Spain 2021 k-NN, Naive Bayes, NNET-neuralnet, SVM, Decision Tree, Random Forest, Logistic Regression AlgorithmAccuracy Sensitivity SpecificityPPV NPV AUC k-NN0.69 0.66 0.71 0.48 0.83 75.48 Naive Bayes 0.75 0.72 0.77 0.56 0.87 80.99 NNET-neuralnet 0.73 0.72 0.73 0.52 0.86 81.73 SVM0.75 0.84 0.71 0.54 0.92 78.49 Decision Tree 0.74 0.66 0.77 0.54 0.85 77.00 Random Forest 0.70 0.66 0.72 0.49 0.84 77.18 Logistic Regression 0.73 0.72 0.73 0.52 0.86 81.37 Miao [ 21 ] China 2024 DL AUC Specificity Sensitivity PPV NPV ConvNeXt0.9090% 90% 96% 77% Cao [ 22 ] China 2025 set model accuracy roc_auc train logistic 0.734 0.815 train rpart 0.734 0.826 train random forest0.749 0.840 train xgboost 0.722 0.804 train svm 0.5870.813 train knn 0.710 0.779 train nnet 0.720 0.790 test logistic 0.692 0.767 test rpart 0.6030.699 test random forest 0.667 0.744 test xgboost 0.641 0.714 test svm 0.590 0.778 test knn 0.679 0.737 test nnet 0.679 0.731 Any discrepancies between reviewers were resolved by the third reviewer. The primary role of the third reviewer is to make the final decision or facilitate discussion to reach consensus in cases where the two primary reviewers disagree. The third reviewer has relevant experience and expertise in the field to ensure the credibility of their decisions. Given the high heterogeneity among the included studies in terms of methodology, AI model types, ultrasound imaging techniques, and diagnostic performance metrics, a quantitative meta-analysis was not feasible. Therefore, a qualitative synthesis was conducted. The results of the included studies were descriptively compared and narratively summarized based on AI model type, diagnostic purpose, and reported performance outcomes (e.g., accuracy, sensitivity, specificity, and error rate) to identify common patterns, strengths, and limitations across studies.

In this systematic review, 808 articles were initially retrieved. After screening titles and abstracts, 23 articles were deemed eligible for full-text assessment. Of these, 18 articles were excluded due to the following reasons: use of non-ultrasound imaging modalities ( n  = 6), insufficient diagnostic performance data ( n  = 4), review or conference papers without original data ( n  = 3), non-English full texts ( n  = 2), and lack of free full-text access ( n  = 3). Ultimately, five articles were included in the systematic review. The methodological quality of the five included studies was assessed based on study design, index test, outcome reporting, and risk of bias. Most studies showed moderate to high quality. Limitations were noted in small sample sizes, single-center datasets, and limited external validation, which may affect generalizability. The risk of bias for each included study was assessed using the QUADAS-2 tool, which evaluates four key domains: patient selection, index test, reference standard, and flow and timing. Each domain was rated as “low,” “high,” or “unclear” risk of bias. Overall, most studies demonstrated a low to moderate risk of bias, indicating acceptable methodological quality. A summary of the quality assessment based on this tool is presented in Table  3 . Table 3 Risk of bias assessment of included studies using the QUADAS-2 tool Study (Author, Year) Patient Selection Index Test Reference Standard Flow and Timing Overall Risk of Bias Applicability Concerns Nouri et al., 2024 [ 18 ] Unclear Low Low Low Low Low Podda et al., 2024 [ 19 ] Low Low Unclear Low Moderate Low Guerriero et al., 2021 [ 20 ] Low Low Low Low Low Low Miao et al., 2024 [ 21 ] Low Unclear Low Low Low Low Cao et al., 2025 [ 22 ] High Low Low High High Moderate Risk of bias assessment of included studies using the QUADAS-2 tool Figure 1 illustrates the PRISMA flow diagram showing the study selection process. Among the included studies, two were conducted in China [ 21 , 22 ], one in Iran [ 18 ], one in Italy [ 19 ], and one in Spain [ 20 ]. The machine learning (ML) algorithms used included SVM, KNN, Random Forest, Logistic Regression, Decision Tree, and ANN. Additionally, convolutional neural networks (CNN) and other deep learning (DL) approaches were employed in some studies [ 18 – 22 ]. All studies reported promising diagnostic performance. However, reporting accuracy alone—often exceeding 0.80—may not fully reflect the true performance of the models, particularly given the relatively low prevalence of endometriosis and unbalanced datasets. Complementary metrics such as sensitivity, specificity, F1-score, and AUC are recommended for a more comprehensive evaluation. Deep learning-based studies achieved the highest reported accuracy values (0.89–0.93 and 90%), largely due to their ability to automatically extract complex spatial patterns from ultrasound images [ 19 , 21 ]. These models required substantial computational resources and benefited from techniques such as transfer learning and data augmentation to mitigate limitations associated with small sample sizes. Machine learning-based approaches relied on manually selected or engineered features. Random Forest and SVM consistently performed well, with reported AUC values around 0.75–0.80 [ 18 , 20 , 22 ]. While their accuracy was slightly lower than deep learning models, they offered better interpretability and required smaller datasets. Deep learning algorithms were more effective for image-based classification and lesion segmentation, whereas traditional ML algorithms were primarily used for predictive modeling and risk estimation. For instance, Podda et al.’s ensemble approach combined multiple CNNs for lesion segmentation, providing a bridge between the two paradigms. Common methodological strengths across studies included cross-validation, feature optimization, and robust model training. Common limitations included small sample sizes, single-center datasets, and limited external validation, which may restrict the generalizability of findings. In summary, both ML and DL approaches demonstrated high diagnostic potential; deep learning excelled in accuracy and complex pattern extraction, whereas machine learning provided interpretability and feasibility for smaller datasets. Future studies should report comprehensive performance metrics, explore hybrid AI models, include larger multicenter datasets, and adopt standardized evaluation frameworks to enhance clinical applicability.

This systematic review was conducted following PRISMA guidelines, with a comprehensive search strategy across multiple databases and dual independent screening of titles, abstracts, and full texts. Only studies explicitly using AI for ultrasound-based diagnosis or prognosis of endometriosis were included, ensuring methodological consistency and relevance. Additionally, the review provides a comparative synthesis of machine learning and deep learning approaches, as well as their reported performance metrics, offering a structured and detailed overview of the current evidence. Several limitations should be acknowledged. First, only freely accessible full-text articles were included, which may introduce selection bias and limit the comprehensiveness of the review. Second, the review included a small number of studies ( n  = 5), which restricts the generalizability of the findings. Third, there was considerable heterogeneity in study designs, AI models, datasets, and outcome reporting, precluding the possibility of performing a meta-analysis. Finally, some studies reported only accuracy without complementary metrics such as AUC, sensitivity, or specificity, limiting the depth of quantitative comparison. Despite these limitations, this review provides a structured, critical synthesis of the available evidence and highlights clear directions for future research.

The findings of this systematic review indicate that artificial intelligence algorithms show the greatest impact in enhancing diagnostic accuracy of endometriosis using ultrasound imaging, while evidence for disease prediction remains more limited. Hybrid and ensemble approaches appear to improve diagnostic performance and could support more informed clinical decision-making. However, due to the small number of studies, heterogeneity in study design, and limited external validation, these findings should be interpreted with caution. Future research should focus on larger, multicenter datasets, standardized evaluation metrics, and the development of interpretable AI models. Additionally, incorporating comprehensive risk factor analysis and multimodal data could enhance both diagnostic and prognostic capabilities, ultimately facilitating personalized patient management and more effective clinical workflows.

This systematic review included five studies that applied artificial intelligence (AI) to ultrasound imaging for the diagnosis or prognosis of endometriosis [ 18 – 22 ]. Five studies focused on diagnosis, while one focused on disease prognosis. Podda et al. employed a multiscale convolutional neural network for automated segmentation of endometriotic lesions, achieving high accuracy and robust AUC [ 19 ]. Miao et al. applied the ConvNeXt algorithm, demonstrating strong performance and the ability to extract complex image features automatically [ 21 ]. Nouri et al. compared SVM, Random Forest, and Gradient Boosting models, reporting moderate-to-high performance (AUC ~ 0.75–0.80) [ 18 ]. Guerriero et al. applied AI models for detecting rectosigmoid endometriosis, achieving results comparable to traditional logistic regression, with faster processing of large datasets [ 20 ]. Cao et al. developed a Random Forest-based predictive model for severe pelvic endometriosis, using SHAP values to provide personalized risk assessment and enhance interpretability [ 22 ]. Although traditional machine learning (ML) models showed reasonable performance, deep learning (DL) models consistently outperformed them. This difference can be attributed to methodological factors: ML relies on manual feature engineering and is limited in capturing complex, non-linear relationships, whereas DL automatically extracts hierarchical features through multi-layer architectures, enabling it to model intricate patterns more effectively. Furthermore, DL benefits more from larger datasets, allowing it to leverage abundant information that ML models may fail to utilize fully. Thus, the superior performance of DL in Table 2 reflects not only the reported metrics but also its inherent ability to learn complex, non-linear features from the data [ 23 ]. Podda et al. utilized a multiscale ensemble approach that further improved the performance of DL models [ 19 ]. DL models offered higher accuracy and automated feature extraction, whereas ML models provided better interpretability and were feasible with smaller datasets. Ensemble approaches enhanced overall diagnostic precision and robustness. Most studies were single-center with relatively small sample sizes, limited external validation, and incomplete reporting of performance metrics. Some studies relied solely on accuracy, which is insufficient for imbalanced datasets. DL models are particularly effective for image classification and lesion segmentation, while ML models are more suitable for risk prediction and prognostic assessment. Hybrid or ensemble strategies combining DL and ML have potential to further improve diagnostic performance. Incorporating diverse, multicenter datasets is crucial to enhance generalizability and clinical applicability. Overall, these Five studies collectively demonstrate the promising potential of AI in improving the diagnosis and prognosis of endometriosis using ultrasound imaging, while also highlighting the importance of methodological rigor, adequate sample size, and comprehensive performance evaluation)18–22(.

Endometriosis is a chronic, non-malignant, estrogen-dependent inflammatory disease characterized by the ectopic growth of endometrial-like tissue outside the uterus, typically associated with symptoms such as pelvic pain, irregular menstruation, and infertility [ 1 – 3 ]. Due to its high prevalence, chronic nature, and significant impact on quality of life, endometriosis is recognized as a major medical condition with considerable socioeconomic burden [ 4 ]. Endometriosis is associated with a high burden of comorbidities, increased healthcare resource utilization, and excess costs, particularly for younger patients whose healthcare needs may differ widely from the older population [ 5 ]. It is estimated that approximately 10% of women of reproductive age are affected, with the prevalence rising to 35–50% among women with pelvic pain or infertility [ 6 ]. Global statistics indicate that approximately 10–15% of women worldwide are affected by endometriosis. In Canada and the United States, the incidence of endometriosis has been reported to range from 5 to 15% among women of reproductive age and from 2 to 5% among postmenopausal women [ 7 ]. In Iran, the estimated prevalence among women of reproductive age ranges from 5% to 20% [ 8 ]. Reports from Germany have shown that 0.05%, 1.93%, and 6.1% of patients were in the 10–14, 15–19, and 20–24-year age groups, respectively. A review of previous studies indicated that global estimates of endometriosis prevalence vary widely, ranging from approximately 2% to 45%, depending on the diagnostic criteria and study population [ 7 ]. Given its high prevalence and substantial disease burden, early and accurate diagnosis of endometriosis is crucial; however, achieving this remains a major clinical challenge. The signs and symptoms of this disease are non-specific and can vary in severity, creating clinical heterogeneity that adds to the diagnostic difficulty. Patients may present with a range of symptoms depending on the type, location, stage, and severity of lesions, including dysmenorrhea, dyspareunia, abdominal pain, chronic pelvic pain, menorrhagia, bowel or urinary symptoms, and subfertility or infertility [ 9 ]. Due to the combination of non-specific symptoms, extensive differential diagnoses, limited provider awareness, unnecessary investigations, and the lack of reliable non-invasive diagnostic tools, many patients experience significant delays before receiving a definitive diagnosis [ 10 ]. The current literature has documented diagnostic delays of up to 6 to 12 years globally. Currently, the gold standard diagnostic procedure for endometriosis remains laparoscopic visualization of lesions followed by histologic confirmation of ectopic endometrial tissue [ 9 ], a costly and invasive process that requires a skilled clinician. Transvaginal ultrasonography is a commonly used clinical technique in endometriosis screening and diagnosis due to its non-invasive nature and widespread availability [ 11 ]. In the past 5 years, the emergence of artificial intelligence (AI) has spread rapidly into healthcare; it has demonstrated marked potential in disease diagnostics, treatments, and a higher-level analysis of large biomedical datasets [ 12 , 13 ]. The introduction of artificial intelligence (AI) in imaging is revolutionizing the diagnosis and management of gynecological disease. AI includes machine learning (ML) models, which are trained to recognize patterns and relationships from input data without explicit programming, and deep learning (DL) models, a subset of ML using artificial neural networks with multiple layers (deep architectures) capable of learning complex representations from data. Convolutional neural networks (CNNs) are a subtype of DL models that can automatically learn spatial hierarchies of characteristics from input images [ 14 ]. CNNs are composed of multiple layers: convolutional layers for extracting image characteristics; pooling layers for reducing dimensionality; and fully connected layers for classification [ 15 ]. ML and DL techniques are often employed in radiomics research to analyze and interpret large quantities of data generated from medical images. Radiomics is a process that extracts features from medical images and provides a quantitative description of the imaging data. The radiomics workflow involves the segmentation of the region of interest (ROI) from the studied image, image processing for subsequent analysis, extraction of radiomics features from the ROI and analysis of the extracted features to identify patterns, correlations and associations with clinical outcomes [ 14 ]. There are several applications of AI in ultrasound imaging, including detection (i.e. the automatic identification of organ structures and other objects of interest), classification (i.e. the analysis of ultrasound images to assess disease status or classify pathology into a specific category) and segmentation (i.e. the delineation of precise lesion boundaries, such as those of ovarian follicles or cysts) [ 16 ]. Considering these advances, the application of artificial intelligence (AI)-based methods to ultrasound imaging represents a novel frontier in the diagnosis and prediction of endometriosis. However, current evidence in this area remains fragmented and inconsistent, particularly regarding model types, diagnostic accuracy, and generalizability. Therefore, the aim of this systematic review is to synthesize the existing literature on the use of AI in ultrasound-based diagnosis and prediction of endometriosis, in order to evaluate the potential, limitations, and future directions of this emerging field.