OP03.02: Artificial intelligence applied to ultrasound diagnosis of pelvic gynecological tumours: a systematic scoping review and meta‐analysis

Pelvic gynecological cancers pose a significant challenge to global health. Ultrasound (US) imaging is crucial for diagnosis, but the subjective interpretation of US images can jeopardise performance. Artificial Intelligence (AI) has the potential to enhance diagnostic accuracy and efficiency. We aimed to evaluate methodologies, performance, and limitations of published AI models for US diagnosis of pelvic gynecological tumours. We conducted a scoping review of studies on PubMed, Embase, Web Of Science, and Cochrane Central, from database inception until August 22, 2022. Pooled estimates of the diagnostic odds ratio (DOR) and area under the summary receiving operating characteristic curve (AUSROC) were calculated. Risk-of-bias assessment was based on an extension of QUADAS-2 with PROBAST and custom items. We conducted subgroup analyses to explore sources of heterogeneity. This study was registered with OSF (https://doi.org/10.17605/OSF.IO/BHKST). 7568 records were retrieved, of which 29 were eligible: 25 on adnexal, three on endometrial, and one on myometrial pathology. All studies were at high risk-of-bias. Only two development studies provided external validation. Pooling of adnexal pathology studies (116 models) yielded an AUSROC of 0.896 (95% CI: 0.851-0.923). AUSROC was of 0.915 (95% CI: 0.779-0.956) for 12 studies considered at ‘very high’ risk-of-bias and 0.867 (95% CI: 0.804-0.912) for eight studies at high risk-of-bias. For the latter eight studies, AUSROC was 0.939 (95% CI: 0.903-0.978) for two deep learning studies vs 0.829 (95% CI: 0.760-0.883) for six machine learning studies. Studies on endometrial and myometrial pathologies were described individually. Current AI models for US diagnosis of gynecological tumours suffer from methodological shortcomings. Even then, AI models for diagnosing adnexal malignancy did not outperform the ADNEX model, the state-of-the-art logistic regression model based on US measurements.