15 May 2024 in IEEE Open Journal of Engineering in Medicine and Biology
Goal: In this study, we demonstrate that a deep neural network (DNN) can be trained to reconstruct highcontrast images, resembling those produced by the multistatic Synthetic Aperture (SA) method using a 128-element array, leveraging pre-beamforming radiofrequency (RF) signals acquired through the monostatic SA approach. Methods: A U-net was trained using 27200 pairs of RF signals, simulated considering a monostatic SA architecture, with their corresponding delay-and-sum beamformed target images in a multistatic 128-element SA configuration. The contrast was assessed on 500 simulated test images of anechoic/hyperechoic targets. The DNN's performance in reconstructing experimental images of a phantom and different in vivo scenarios was tested too. Results: The DNN, compared to the simple monostatic SA approach used to acquire pre-beamforming signals, generated better-quality images with higher contrast and reduced noise/artifacts. Conclusions: The obtained results suggest the potential for the development of a single-channel setup, simultaneously providing good-quality images and reducing hardware complexity
5 April 2025 in Computers in Biology and Medicine
Sub-optimal uterine-placental perfusion and fetal nutrition can lead to intrauterine growth restriction (IUGR), also called fetal growth restriction (FGR). Antenatal cardiotocography (CTG) can aid in the early detection of IUGR. Reliably diagnosing IUGR before delivery remains challenging, and deep learning (DL) techniques offer potential solutions. This paper describes the development of a DL approach to predict an IUGR condition at birth by using CTG signals collected during antenatal monitoring. Our method is encapsulated in the concept of a two-step training process of a ResNet architecture. The primary focus is on the minimization of data loss, which motivates the division into "presumed" and "confirmed" datasets, which is employed to distinguish based on the presence of information at birth. The method involves fine-tuning: the initial training utilizes "presumed" data to train the network, and the subsequent training employs data representing certain knowledge to refine its performance. The DL model reaches a balanced accuracy of 80% on a hold-out test set of confirmed cases, which is better than what obtained by using standard clinical guidelines. The results of our work are compared to the results of similar papers dealing with the prediction of IUGR condition at birth and in general with the prediction of fetal pathological conditions. Our final results are obtained using a very large dataset compared to other papers reported in the literature. The inclusion of DL methods on CTG signals may complement imaging technologies and improve the early detection of IUGR.