OC15.01: Radiomics analysis of early pregnancy ultrasound images to predict viability at the end of first trimester

Identify radiomic ultrasound features from the pregnancy of unknown viability diagnosis images, which may predict subsequent loss. Radiomics offers a quantitative method to assess heterogeneity between images, which can be more stable compared to subjective assessment. A retrospective, multi-site study, included 500 cases of early pregnancy ‘pregnancy of unknown viability’ (PUV). Longitudinal cross-sectional ultrasound images were extracted and segmented 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). Clinical data including age, ethnicity, gestational age, symptoms and operator-recorded ultrasound features were collected. A radiomics prediction model was developed by comparing several ML models using convolutional neural network-based models and traditional radiomics features. The area under the ROC curve (AUC), F1-score, and recall to assess model performance. The QCCH and SMH data sets were in the majority well matched and consisted of 53.3% and 53% miscarriage cases by end of first trimester, respectively. The best performing radiomics classification model (Support Vector Machine model in combination with Mutual Info feature selection technique) for predicting miscarriage from early pregnancy PUV ultrasound images; achieved a recall of 0.81 and an AUC of 0.68 (F1-score 0.67), 0.71 (F1-score 0.70) and 0.72 (F1-score 0.67) in the QCCH training, QCCH validation and SMH test set respectively. This machine learning method is the first to use radiomics to predict early pregnancy outcomes, and has been validated and externally tested. This study sets the stage for future trials to prospectively evaluate the performance of the model, which can then be used to help patients navigate the uncertainty of a PUV early pregnancy diagnosis.