Search for a command to run...
To determine whether there are radiomic ultrasound features of early pregnancy when viability is unknown, which in combination with clinical features, may predict subsequent loss. Multi-centre retrospective cohort study, which included 500 cases of pregnancies of unknown viability (PUV) collected from January 2021 to January 2023. Longitudinal ultrasound images were identified from Queen Charlotte's and Chelsea Hospital (QCCH), London (n = 400, split 8:2 for training and validation) and St Mary's Hospital (SMH), London (test data set n = 100). Images were extracted and segmented to include firstly the gestation sac and secondly the sac endometrial border. A segmentation model was developed using a deep learning (DL) model (multi-task nnUNet v2) and standard Dice Coefficient (DICE) was used to measure performance. A prediction model, using clinical and radiomic features, was developed by comparing several machine learning (ML) methods. The area under the ROC curve (AUC), F1-score, and recall were used to assess model performance. The QCCH and SMH data sets were in the majority well matched and consisted of 53.3% and 53.0% miscarriage cases by the end of first trimester, respectively. The DL segmentation model for gestation sac achieved a mean DICE score of 0.950 and 0.940 in the training and test data sets respectively. The segmentation model for the sac endometrial border achieved a mean DICE score of 0.917 (QCCH) and 0.922 (SMH). The best performing PUV outcome classification model (XGBoost and LASSO) for predicting miscarriage (PUVPS model); achieved an AUC of 1.00 (F1-score 1.00), 0.92 (F1-score 0.79) and 0.84 (F1-score 0.76) in the QCCH training, QCCH validation and SMH test set respectively. We have developed an end-to-end radiomics-based model to segment and predict early pregnancy outcomes. The main limitation of this study is its sample size, which can make a ML model prone to overfitting. This study sets the stage for future trials to prospectively evaluate the performance of the PUVPS model, in a large multi-centre cohort, which can then be used to help patients navigate the uncertainty of a PUV early pregnancy classification.