OP02.09: AIRFRAME: artificial intelligence for recognition of fetal brain anomalies from ultrasound images of the first trimester

Objectives Abnormal sonographic findings of the posterior cranial fossa (PCF) at first trimester fetal brain midsagittal view (abnormal ratio between the length of the brainstem (BS) and its distance from occipital bone (OB)) are early markers of open spina bifida (OSB) or Dandy Walker malformation (DWM). Our study aims to develop an artificial intelligence (AI) algorithm for the automatic classification of PCF ultrasound images as normal or abnormal at 11-14 weeks of gestation. Methods AIRFRAME is a multicentre study involving 16 referral centres. In this retrospective phase, 251 midsagittal images of the fetal brain at 11-14 weeks were used to develop the algorithm; 150 cases were classified as normal and 101 as abnormal (OSB n = 42, DWM n = 59) after diagnostic confirmation. The region of interest was selected visualising the three hypoechoic areas of the PCF, and quantifying it via BS/BSOB measurement. Each image was manually segmented and optimised for deep learning (DL) applications. The dataset was randomly split into 70% learning (n = 175) and 30% test (n = 76) sets. A threefold cross-validation on the learning set was conducted to train different neural networks and select the optimal model. The test set was exclusively used to evaluate classification performance of the selected model. To combine the multiple models obtained, individual predictions probability were averaged. Results The customised convolutional neural network demonstrated an excellent performance on the test set in differentiating normal vs abnormal cases (receiver operating characteristic area under the curve = 0.81, sensitivity = 0.74, specificity = 0.69). Conclusions The preliminary results of our study demonstrate that a DL algorithm has the potential to support clinicians in the sonographic evaluation of fetal PCF at 11-14 weeks of gestation through a fully automatic data processing, potentially increasing the performance of first trimester ultrasound in detecting major CNS anomalies. The prospective phase will assess its clinical applicability into the routine practice.