Artificial intelligence shows promise for ultrasound-based diagnosis of endometriosis

Endometriosis remains one of the most challenging gynecologic conditions to diagnose, often requiring years before confirmation. In a new systematic review published in Reproductive Health, investigators evaluated whether artificial intelligence (AI) applied to ultrasound imaging could improve diagnostic accuracy and prediction of endometriosis, addressing a longstanding gap in noninvasive detection.

Endometriosis affects an estimated 10% to 15% of women worldwide and up to 35% to 50% of those presenting with pelvic pain or infertility. Despite its prevalence, diagnosis is frequently delayed by 6 to 12 years, largely due to nonspecific symptoms, variable disease presentation, and reliance on laparoscopy with histologic confirmation as the diagnostic gold standard. Transvaginal ultrasound is widely used in clinical practice, but its sensitivity depends heavily on operator expertise and disease stage.

Study design and methods of the systematic review

To assess whether AI could enhance ultrasound-based detection, the authors conducted a systematic review in accordance with PRISMA 2020 guidelines. Searches of PubMed, Scopus, Web of Science, and Google Scholar were completed through April 30, 2025, using predefined keywords related to artificial intelligence, diagnosis, prediction, endometriosis, and ultrasonography.

Only English-language, peer-reviewed original studies evaluating AI algorithms applied to ultrasound images were included. Studies were required to report diagnostic performance metrics such as accuracy, sensitivity, specificity, or AUC. Methodological quality and risk of bias were assessed using the QUADAS-2 tool.

Overview of included studies and AI models

Out of 808 screened records, 5 studies met inclusion criteria. These studies were conducted in Iran, Italy, Spain, and China and collectively evaluated both machine learning (ML) and deep learning (DL) approaches for diagnosing or predicting endometriosis using ultrasound imaging. Study sample sizes ranged from 53 images to more than 500 patients.

The AI techniques included traditional machine learning models—such as support vector machines, random forest models, logistic regression, and k-nearest neighbors—as well as deep learning approaches using convolutional neural networks.

Diagnostic performance of deep learning models

Across studies, deep learning models demonstrated the strongest diagnostic performance. Reported accuracy values ranged from 0.89 to 0.93, with AUC values around 0.90. Sensitivity ranged from 0.78 to 0.92, and specificity ranged from 0.74 to 0.89.

These models were particularly effective for image-based classification tasks and lesion segmentation, including differentiation between ovarian endometriomas and benign cystic lesions and detection of deep infiltrating endometriosis. Their ability to automatically extract complex image features contributed to higher overall diagnostic accuracy.

Machine learning approaches offer interpretability with modest accuracy

Machine learning approaches also showed diagnostic utility, though with slightly lower performance metrics. ML model accuracy generally ranged from 0.80 to 0.85, with AUC values between 0.75 and 0.80.

While less accurate than deep learning models, machine learning approaches required smaller datasets and offered greater interpretability. Some studies used explainability techniques to identify contributing clinical or imaging features, which may support clinician understanding and trust.

Study limitations and risk of bias assessment

Methodological quality across studies was generally moderate to high. However, common limitations included small sample sizes, single-center datasets, limited external validation, and inconsistent reporting of diagnostic performance metrics.

The authors noted that reliance on accuracy alone may overestimate performance, particularly in datasets with low disease prevalence. Because of heterogeneity in study designs, AI models, and ultrasound techniques, quantitative meta-analysis was not feasible.

Clinical implications and future research directions

The review highlights artificial intelligence as a promising adjunct to ultrasound imaging for endometriosis diagnosis. AI-assisted tools may help improve diagnostic efficiency, reduce dependence on invasive procedures, and support earlier detection.

However, the authors emphasized that clinical implementation remains premature. Future research should focus on larger, multicenter datasets, standardized evaluation frameworks, and externally validated, interpretable AI models to support integration into routine gynecologic practice.

Reference

Esmailzadeh, A., Rashki Kemmak, A., Sezavar Dokhtfaroughi, S. et al. Investigating the role of artificial intelligence in the diagnosis and prediction of endometriosis using ultrasound images: a systematic review. Reprod Health (2026). https://doi.org/10.1186/s12978-025-02245-1