OC07.06: Artificial intelligence for automated detection of congenital brain anomalies in the first trimester: the Rotterdam Periconception Cohort
Year: 2024
Type: conference-abstract
Abstract: Congenital brain anomalies, such as anencephaly, occur with an incidence of approximately 0.1% and often necessitate termination of pregnancy due to their severity. While most of these anomalies are present in the first trimester, the detection rate is low (53% to 73%). To improve this rate, we aim to develop an artificial intelligence (AI) algorithm for automatic detection of congenital brain anomalies, using first trimester three-dimensional (3D) ultrasound images. We utilised 3D ultrasound images acquired in the 9 gestational week of pregnancy from the Rotterdam Periconception Cohort. The dataset contains images of 404 embryos, known to have developed normally, and 8 images of 4 embryos with a visible brain anomaly. 3D ultrasound datasets of normally developing embryos were used to train an AI algorithm for anomaly detection. This algorithm is an extended version of a state-of-the-art two-dimensional anomaly detection algorithm called EfficientAD, which learns structural features typical for normal embryos, after which it can flag the atypical cases. The trained algorithm was tested on previously unseen data of 41 normal and 8 abnormal datasets. We computed the area under the receiver operating characteristic curve (AUROC), accuracy, precision, and recall. The AUROC value for detecting abnormal ultrasound datasets was 90%. For the optimal threshold, we had an accuracy of 80%, a precision of 44% and a recall of 100%. Here, we showed the first steps towards automatic detection of brain anomalies in first trimester pregnancies using 3D ultrasound images. The next step is to evaluate if the abnormal features correspond with the brain anomalies. In the future, we will extend this algorithm towards a broader age range and towards all anatomical structures to enable automated congenital anomaly screening during the first trimester.
DOI: 10.1002/uog.27767
Language: English
Field: Medicine
Domain: Health Sciences
Open Access status: bronze
OpenAlex ID: W4402359401