Automatic Recognition on Fetal Pulsed-Wave Doppler Envelope using Neural Networks

Pulsed-Wave Doppler (PWD) echocardiography is one of the standard techniques for antenatal cardiological diagnosis. When applied to fetuses, this technique is challenging since, beyond being intrinsically operator-dependent, different issues related to the fetal heart size, the fetal movements and the ultrasound artifacts appear. In long PWD recordings, the signal segments completely meaningful for a morphological analysis are then limited in number and duration. In this work, a neural network-based approach for the automatic identification of the fetal beats showing the most important waves of the PWD is presented and evaluated on real signals. The proposed algorithm works on a couple of 1D signals, representing the PWD envelope extracted from the video. For the validation, a small dataset was created, including 8 records from four voluntary pregnant women (21st to 27th gestational week), 10 seconds long each. An expert cardiologist annotated the dataset. The performance of the method was evaluated through a 4-fold cross-validation scheme, revealing an average accuracy up to 87.8%. This confirms the validity of the proposed approach, laying the basis for future improvements.