(697) Identifying signs of congenital infections with Artificial Intelligence

Periventricular and subependymal cysts are frequent sonographic findings, often overlooked but which may be one of the rare manifestations of Cytomegalovirus (CMV) infection. An Artificial Intelligence (AI) able to automatically recognize these findings could serve as a safety net, prompting practitioners to conduct further investigations when necessary.

Study Design:
This study evaluated an AI model (not yet approved) for detecting subtle cerebral findings, such as images of periventricular (PVC) or subependymal cysts (SC) and periventricular (PVN) or subependymal nodules (SN). Trained on 13797 images from 74 centers, including 1180 images of cysts or nodules from 70 CMV-seroconverted pregnancies and 143 fetuses with isolated brain cyst. The model was evaluated on a test set from a center entirely excluded from the training set where experts identified 84 pathological axial brain images (59 PVC, 8 SC, 12 PVN, 5 SN) and 159 pathological coronal brain images (81 PVC, 21 SC, 19 PVN, 38 SN). The specificity was evaluated on 118 normal axial brain images and 14 coronal brain images from this test center, adding 126 normal coronal brain images not used for training but coming from the 74 training centers.

Results:

The overall sensitivity was 84 %. The sensitivity was notably higher on coronal view 86,2%, than on axial views 78,6%. The system also exhibits a higher sensitivity on SC, 93,1%, or SN (79,1%) compared to PVC 86,4% or PVN (71%) where ultrasound artefacts are frequent on the anterior horns of the ventricle. The specificity of the system was 94,1% on axial brain images, 97,1% on coronal brain images.

Conclusion:
AI has the potential to improve ultrasound screening for congenital brain cysts and, consequently, the detection of certain congenital infections. Further work is needed to strengthen the sensitivity of the model on nodules and on periventricular cysts of the anterior horn of the ventricle in axial views without significantly compromising specificity. In addition, learning the other signs of congenital infection should enable AI to greatly improve the relevance of the highlighted fetus.