Prediction of uterine cavity conception environment using two-dimensional transvaginal ultrasound imaging semantic feature-based machine learning: a case-control study.

backgroundIndependently investigating the association between pregnancy outcomes and the uterine cavity conception environment (UCCE) is challenging. Therefore, this study aimed to employ a range of machine learning algorithms to systematically analyze the semantic features derived from two-dimensional transvaginal ultrasound (TVUS) images that characterize the UCCE.

methodsThis case-control study was conducted at the Women and Children's Hospital of Chongqing Medical University. Preoperative TVUS findings in patients diagnosed with infertility who subsequently achieved clinical pregnancy within 1 year following hysteroscopic intervention alone were classified as abnormal UCCE (case group), whereas TVUS findings in the uterine cavity of women who successfully achieved clinical pregnancy through follicular monitoring during the same cycle were classified as normal UCCE (control group). We randomly divided 442 participants (cases, n = 244; controls, n = 198) into the training (n = 309) and testing (n = 133) sets. A structured description of the ultrasonic semantic features of UCCE was derived by integrating the images and diagnostic reports. To employ logistic regression (LR), the Boruta algorithm, Least Absolute Shrinkage and Selection Operator for univariate feature selection, and five machine learning models-LR, random forest (RF), XGBoost, support vector machines (SVM), and LightGBM-were constructed, followed by a comprehensive evaluation of their predictive performance. Additionally, the Shapley additive explanation technology was employed to elucidate the output results from the best-performing model. A nomogram was created for diagnostic prediction model visualization.

resultsEndometrial thickness, echogenicity, polyps, polyp size, continuity and uterine morphology, and septum type were the semantic features with the highest UCCE predictive value. In the testing set, XGBoost and LightGBM showed the highest (0.982) and lowest (0.902) area under the curve (AUC), respectively. XGBoost, SVM, RF, and MLR attained excellent prediction accuracy (0 < Brier score < 0.1) across the training and testing sets. Decision-curve analysis demonstrated that all models provided substantial net clinical benefits. A significant difference was observed in the nomoscore between the normal and abnormal UCCE groups (p < .001).

conclusionsMachine learning models using semantic features derived from two-dimensional TVUS images enabled accurate preliminary screening for natural conception preparation or before embryo implantation.